Adhesion between neighboring grains can make instances difficult to distinguish, while the scale of a plant’s point cloud creates a substantial segmentation problem. PRSNeRF addresses the task in two stages: CoNeRF produces a three-dimensional representation from multi-view images, and PRSNet assigns separate identities to individual grains. The researchers tested reconstruction on the Blur Rice benchmark and assessed segmentation on the RiceNeRF-IS dataset.

CoNeRF integrates image deblurring, contour-guided sampling and a rendering-projection consistency mechanism. These components are designed to improve geometric fidelity and suppress floating artifacts in the reconstruction. Under a fixed budget of 5,000 training iterations, the method reduced the scaled NeRF-space Chamfer distance from 10.61 to 7.81.

It also required substantially less time than the 30,000-iteration NeRFacto baseline: training time decreased from 368 seconds to 77 seconds, and point-cloud export time fell from 120 seconds to 10 seconds. These results apply to the benchmark conditions and comparison setup used in the study. PRSNet handles the next stage by identifying separate grain instances in the reconstructed point clouds.

Its architecture includes a Mamba-based encoder, adaptive Gaussian detail enhancement, multi-scale feature fusion and three-dimensional dynamic anchor boxes. In plant-level five-fold cross-validation on RiceNeRF-IS, PRSNet achieved an overall average precision of 87.31%, outperforming OneFormer3D by 22.30 percentage points. The segmented point clouds then provided the basis for automated measurements of grain number, grain dimensions and voxel-based occupancy.

The study presents PRSNeRF as a connected workflow linking image-based reconstruction, point-cloud instance separation and quantitative trait extraction, with a reported end-to-end processing time of approximately 3.3 minutes per plant. Its demonstrated performance is limited to controlled indoor acquisition and the specific datasets and baselines evaluated. The evidence does not establish equivalent results for field-grown rice, outdoor imaging or other crop species, and the early-access article is described as subject to further editorial revision.

Within the tested setting, however, the workflow shows that improvements in three-dimensional rendering and point-cloud analysis can support faster automated rice phenotyping while producing grain-level measurements.

Subject of Research: Automated three-dimensional rice-grain phenotyping using NeRF reconstruction and Mamba-based point-cloud segmentation

Article Title: A 3D grain-level phenotypic extraction framework for rice using contour-guided NeRF reconstruction and mamba-based point cloud segmentation

Article References: Qu, F., Huang, H., Song, W., Zhang, W., Wu, L., Tang, C., Zhang, J., & Sun, Y. (2026). A 3D grain-level phenotypic extraction framework for rice using contour-guided NeRF reconstruction and mamba-based point cloud segmentation. Plant Methods. https://doi.org/10.1186/s13007-026-01603-9

Image Credits: AI Generated

DOI: 10.1186/s13007-026-01603-9

Keywords: rice phenotyping, 3D reconstruction, neural radiance fields, point-cloud segmentation, Mamba networks, computer vision, grain traits, rice panicles, plant breeding, grain-level, phenotypic, extraction

The late-season experiment was conducted at the Gulf Coast Research and Education Center in Wimauma, Florida. Researchers selected four plots located on two raised beds, with each plot containing a different genotype for every plant. The four plots held 16, 18, 14 and 16 valid plants, respectively, for a total of 64. Ground and aerial imaging took place on March 5 and March 12, 2024, with each session covering two plots. Two days later, the plants were harvested for ground-truth measurements. Runners, fruits, flowers, crowns and roots were removed, leaving leaves and stems. The retained material was weighed fresh, dried in an oven for five to seven days and weighed again after reaching constant weight. Dry biomass was the outcome used in the analysis because fresh mass can vary with temperature, weather and irrigation timing.

Two imaging platforms supplied the observations. A custom ground vehicle carried three Canon cameras: one top-facing Canon R8 and two side-facing Canon 7D cameras. An RTK global navigation satellite system recorded the vehicle’s trajectory, while a dark tarp and two artificial light sources were added to reduce glare, inadequate illumination and wind-related movement. The cameras were synchronized and triggered every second as the vehicle traveled at 0.13 metres per second, producing substantial image overlap. This close-range platform captured finer leaf detail than the aerial system, but its estimated camera positions followed three linear trajectories around the plants. That limited angular coverage was sufficient for SfM photogrammetry but not for NeRF or 3DGS with SuGaR: attempts to reconstruct recognizable plant structure from the ground images with those two methods were unsuccessful.

The UAV therefore provided the imagery for all three reconstruction approaches tested there. Researchers flew a DJI Mavic 3 Pro in point-of-interest mode, directing it around a central target while the gimbal kept the plants centered. Each approximately 36-second recording was captured from 12 metres above ground at 60 frames per second and 3840 by 2160 pixels. The gimbal was tilted 25 degrees below horizontal and used threefold optical zoom. Extracting frames at five per second produced about 180 images per plot. Unlike a conventional linear mapping flight dominated by top views, the circling trajectory provided overlapping perspectives around the canopy. Surveyed plot- and field-level ground control points were used as georeferencing references for the resulting models.

SfM, NeRF and 3DGS with SuGaR shared the camera-alignment stage in Agisoft Metashape. SfM used the aligned images to generate dense point clouds directly. The other workflows used the same images and camera positions in COLMAP format. The Nerfacto implementation in Nerfstudio was trained for 30,000 iterations, with 90% of frames assigned to training and 10% held out to assess rendering quality through peak signal-to-noise ratio, structural similarity and learned perceptual image patch similarity. The input frames were automatically downscaled to 1600 by 900 pixels for NeRF training. The 3DGS and SuGaR workflow used density regularization, a high-resolution mesh and a high-poly configuration refined over 7,000 iterations. Its output was an explicit surface mesh, whereas standard 3D Gaussian Splatting represents a scene as discrete Gaussian primitives without a defined surface between them.

Individual plants were separated manually from the plot-level models. An RGB hue filter removed mature fruits, flowers and black mulch from the point clouds, while the Gaussian-splatting meshes were manually denoised to remove the same non-canopy material. Researchers then fitted a three-dimensional convex hull around each retained leaf-and-stem structure and used its enclosed volume as the predictor of dry biomass. The convex hull follows an outer envelope rather than the true internal geometry of a strawberry plant, so it can include gaps among drooping leaves, curled leaf margins and stems. Incomplete camera coverage, especially in shadows, can reduce the reconstructed extent in the opposite direction. Because the study did not independently measure true canopy volume, it could not quantify the net size or direction of these competing errors.

Canopy volumes were strongly correlated among the four reconstruction methods, with pairwise R² values from 0.93 to 0.98, but correlation did not imply perfect agreement. Mean bias differed significantly from zero for every method pair, and relative root mean square error ranged from 11.6% to 25.1% of mean volume. UAV SfM produced volumes that were 15.2% lower on average than ground SfM, and the difference increased with volume. UAV NeRF volumes were 20.8% higher on average than UAV SfM volumes. The closest agreement occurred between UAV SfM and UAV 3DGS with SuGaR: their mean bias was 3.9%, relative RMSE was 11.6%, and Lin’s concordance correlation coefficient was 0.98. The methods shared ground control points and the same convex-hull volume procedure, so their consistency cannot independently validate the georeferencing correction.

Linear regression converted each canopy-volume estimate into predicted dry biomass, with canopy volume as the sole predictor. Accuracy was assessed by plant-level five-fold cross-validation repeated with 100 random fold assignments, so every plant was predicted once by a model that had not been trained on it. UAV 3DGS with SuGaR had the highest mean performance, with R² of 0.85, RMSE of 9.32 grams and mean absolute percentage error of 16.02%. UAV SfM followed with R² of 0.83, RMSE of 9.91 grams and MAPE of 16.08%. Ground SfM achieved R² of 0.79, RMSE of 11.09 grams and MAPE of 16.68%, while UAV NeRF achieved R² of 0.77, RMSE of 11.07 grams and MAPE of 17.46%. A paired bootstrap analysis of all six method pairs found no RMSE difference that remained significant after Holm adjustment for six comparisons. The non-significant result did not demonstrate equivalence, and the confidence intervals left differences of roughly two grams around the observed estimates unexcluded.

Two plants had markedly greater biomass than the rest: one Felicity plant with approximately 115 grams of dry biomass and one Camarosa plant with approximately 145 grams. Only the Camarosa observation was influential under every method, with a Cook’s distance far above the study threshold, and its measured biomass exceeded the value predicted from canopy volume. The similarly large Felicity plant was predicted accurately and had low influence. The researchers proposed that denser foliage and more leaf and stem mass per unit of enclosed canopy volume might explain the Camarosa result, but foliage density was not measured and the explanation was based on a single plant. It remains a hypothesis rather than a demonstrated mechanism. More generally, the external convex-hull measurement does not directly capture genotype-dependent internal structure, which could contribute residual variation to the volume–biomass relationship.

Timing measurements showed that fast data acquisition did not yet translate into a fully automated high-throughput phenotyping pipeline. A UAV recorded a complete plot in about 36 seconds, while the total automated UAV SfM workflow required approximately 12.1 minutes per plot. UAV NeRF required 11.6 minutes, ground SfM 11 minutes, and UAV 3DGS with SuGaR 54.6 minutes. Destructive biomass sampling required about 10 minutes per plot, followed by five to seven days of drying. These comparisons excluded manual plant segmentation, point-cloud or mesh denoising, ground-control-point placement and platform setup. The demonstrated advantages of the UAV workflow were therefore rapid, non-destructive and repeatable acquisition, rather than complete end-to-end speed. Automating segmentation and marker placement, and reducing the computational burden of 3DGS with SuGaR, were identified as prerequisites for scaling the method to large trials.

The study has several limits that constrain generalization. It included only late-season plants from one experimental site, so performance at earlier growth stages, in other fields or seasons, or in other crops remains unknown. Earlier canopies may be less occluded but contain proportionally larger internal gaps, creating a different balance between reconstruction error and convex-hull inflation. Plants within a plot also share a reconstruction, ground-control-point registration and lighting conditions, so plant-level cross-validation may overstate performance for entirely new plots. Leave-one-plot-out validation gave similar results, but with only four plots those estimates remain imprecise. Preprocessing was not identical across methods: point clouds underwent hue filtering, whereas meshes were manually denoised, and operator repeatability was not assessed. The field experiment therefore supports the feasibility of non-destructive strawberry biomass estimation, while additional seasons, independent plots, common automated preprocessing and independently measured canopy volume are needed to establish how well the approach generalizes.

Subject of Research: Comparison of UAV and ground-based 3D reconstruction methods for estimating dry biomass in field-grown strawberries

Article Title: UAV-based 3D canopy reconstruction for non-destructive strawberry biomass estimation: Comparison of SfM, NeRF, and Gaussian Splatting

Article References: Shen, K., Wang, X., Contreras, A. B., Ji, L., Dalid, C., Zheng, C., Smith, K. E., Li, C., & Whitaker, V. (2026). UAV-based 3D canopy reconstruction for non-destructive strawberry biomass estimation: Comparison of SfM, NeRF, and Gaussian Splatting. Smart Agricultural Technology, 15, Article 102649. https://doi.org/10.1016/j.atech.2026.102649

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102649

Keywords: strawberry biomass, UAV imaging, 3D reconstruction, SfM photogrammetry, Neural Radiance Fields, 3D Gaussian Splatting, plant phenotyping, canopy volume, dry biomass, field phenotyping, UAV-based, canopy

The specimen, catalogued as SMNK-MAM 3924 at the State Museum of Natural History Karlsruhe, arrived in 1977 among numerous zoological, palaeontological and geological objects from the collections of the former State Museum of Natural History in Mannheim. Its original label is missing, but a fading ink inscription on the top of the skull reads Chaeropus castanotis, a junior synonym of the southern pig-footed bandicoot (Chaeropus ecaudatus). A second, nearly illegible inscription on the right temporal region shows the number 95 preceded by a short sequence of letters that may read Mai, the German word for May, suggesting either a collection or receipt date of May 1895.

Had the identification been correct, the specimen would have been extraordinarily significant. According to a 2019 review cited in the study, only ten specimens of the southern pig-footed bandicoot are held in Australian museums, with seven others in the Natural History Museum in London and one in the American Museum of Natural History in New York, and nowhere else. The recently described northern pig-footed bandicoot (Chaeropus yirratji) is similarly scarce, known from just fifteen Australian specimens plus single holdings in London, Cambridge and Paris. Most of these specimens were collected during a handful of expeditions between 1838 and 1907, before both species vanished from the semi-arid zones of New South Wales, Victoria, South Australia and Western Australia.

The problem facing curators in Karlsruhe was that the skull had disintegrated into dissociated fragments, most of them fragile, which hampered both morphological evaluation and any attempt at physical reconstruction. Gluing such delicate pieces back together risked further damage and would have permanently altered the specimen. The team instead chose a fully digital route: scan every fragment at high resolution, clean the resulting data, and reassemble the skull on a computer screen.

The scanning was carried out with an Artec Space Spider, a handheld portable 3D scanner whose manufacturer specifications promise accuracy of up to 0.05 millimetres and a resolution of up to 0.1 millimetres. The device weighs just 0.85 kilograms, captures both geometry and surface colour, and operates target-free at 7.5 frames per second on objects as small as five millimetres. Each fragment was placed on a rotating platform to allow a full 360-degree view and scanned from two or three different angles, covering dorsal and ventral or lateral and medial sides. For the two thin, mediolaterally compressed halves of the lower jaw, the researchers borrowed a so-called third hand, a jig normally used in soldering, to hold the elements steady and capture their delicate edges. The scanner’s highest sensitivity setting was used throughout.

Post-processing was performed in ARTEC STUDIO 16 PROFESSIONAL. Unwanted elements such as images of the scanning platform and assorted artefacts were removed manually, and the two to three captures of each fragment were merged into a single 3D image aligned with manually set markers. Incomplete areas of bony surface and residual artefacts were repaired with the software’s lasso selection and fix-holes tools before textures were applied and the files exported in stl, mtl and obj formats. The final assembly took place in Blender 4.1, an open-source 3D modelling package, where published photographs of the skulls of Chaeropus ecaudatus and Chaeropus yirratji served as reference material alongside the actual specimen. The finished renderings were exported as tif files at a resolution of 10,024 by 9,077 pixels.

Six fragments proved sufficient to align into an almost complete skull, with no missing elements added and no parts mirrored. Not everything went smoothly. Attempts to scan the isolated teeth failed entirely, as the scanner’s resolution was not up to the task. The flat, thin dentary halves were prone to bubble-like artefacts on the bone surface that the software could not automatically distinguish from real anatomy, so these had to be removed manually and the areas reconstructed using the surrounding surface as a template. Even at high resolution, slight structural differences between the 3D images and the original fragments required manual correction, though the authors note that none of these adjustments touched any diagnostic characters relevant to species identification. The entire workflow, from scanning through adjustment to virtual reconstruction, took multiple days.

The reconstructed skull told a clear story, and it was not the story on the label. The snout is long and wide, whereas Chaeropus species have short, narrow snouts. There is no accessory palatal fenestra, while Chaeropus always carries a pair of lateral and medial accessory palatal fenestrae. In lateral view the ventral edge of the maxilla is almost straight, rather than showing the deep curvature that elevates the canine and premolars well above the molars in pig-footed bandicoots. The antorbital fossa is weakly developed instead of deep, the zygomatic arch is thin in ventral view rather than thick, and the dentary is very long rather than short, lacking the triangular process at the base of the ascending ramus that expands the dentary laterally in Chaeropus. Features pointing to Perameles include a nasal that terminates well anterior to the anterior margin of the orbit, the presence of the primary foramen ovale, and an alisphenoid tympanic process that is not hypertrophied. The absence of accessory fenestrae, the weak antorbital fossa and the short diastema between the canine and first premolar narrow the field to two species, the southern long-nosed bandicoot (Perameles nasuta) and the northern long-nosed bandicoot (Perameles pallescens). A poorly expanded submasseteric crest on the dentary tipped the balance toward Perameles nasuta, a species that remains common in eastern Australia and even dwells in suburban gardens.

The authors emphasise that the method is remarkably efficient. Without extensive prior training in the hardware and software, scanning took only a few hours, post-processing was completed within a working day, and the virtual reconstruction required a couple of days. One practical limitation emerged: dimensional information could not be transferred from ARTEC STUDIO 16 PROFESSIONAL to Blender through the obj files, so a scale bar had to be adjusted manually using a photograph of the isolated fragments with a scale bar as a template. Dimensional accuracy was then verified by taking real measurements of the left dentary and comparing them with its virtual counterpart. Because the resulting images can be shared online and compared directly with specimens in other museums, virtual reconstruction clearly facilitates identification to species level, a considerable advantage for collections holding damaged or disarticulated material.

How the skull actually reached Mannheim remains an open question. Georg Arnold, who served as curator of the State Museum of Natural History Mannheim from 1876 until his death in 1896, was in contact with the German botanist Ferdinand von Müller in Melbourne, and Arnold’s son Ernst travelled to Australia and collected insects around Melbourne in the 1890s. The researchers consider it not unlikely that either von Müller or Ernst Arnold forwarded the specimen to Mannheim, a hypothesis that fits the faded May 1895 inscription, although no corresponding record exists. The surviving museum catalogues list only very few Australian marsupials and none from Peramelemorphia at all. The latest catalogue entries date from 1877 for mounted mammals and 1901 for skeletons and skulls, yet the museum demonstrably received further specimens at least into the 1930s, including the vast ethnological, palaeontological and zoological collections of Gabriel von Max acquired in 1917. Who misidentified the skull as Chaeropus ecaudatus, and how it entered the collection, therefore remains obscure, a small mystery preserved in fragile bone and fading ink, now at least bearing the right name.

Subject of Research: Virtual 3D reconstruction of a fragmented bandicoot skull to correct its species identification

Article Title: Virtual 3D-reconstruction facilitates identification of fragmented peramelemorph skull

Article References: Specht, J., Manegold, A., & Travouillon, K. J. (2026). Virtual 3D-reconstruction facilitates identification of fragmented peramelemorph skull. Natural History Collections and Museomics, 3, Article e192533. https://doi.org/10.3897/nhcm.3.192533

Image Credits: AI Generated

DOI: 10.3897/nhcm.3.192533

Keywords: bandicoot, 3D reconstruction, museum collections, Peramelemorphia, Chaeropus, Perameles nasuta, surface scanning, virtual reconstruction, marsupials, taxonomy, natural history, specimen identification

The system, called ABox-PC, was developed by a research team working at a vegetable experimental station in Yan’an, Shaanxi Province, China. The team scanned tomato plants of the cultivar Jinpeng101 using a high-precision time-of-flight RGB-D camera capable of millimeter-level depth sensing across a range of 0.25 to 5.46 meters. Each scan file contained roughly 800,000 points describing the scene. After harvesting, the researchers manually measured each fruit’s horizontal diameter, vertical diameter, volume, and weight three times, averaging the values to create ground-truth references. In total, 222 fruits were collected, each annotated at the fruit level and paired with manual phenotypic measurements, forming the first such systematically paired dataset at the station.

The core problem the researchers set out to solve is geometric rather than photographic. A tomato fruit is a compact, approximately closed organ whose key traits, including its horizontal diameter, vertical diameter, shape ratio, and volume, are determined by its external contour and surface curvature. When occlusion removes part of the observable contour, the estimated centroid shifts, the principal orientation changes, and the measured spatial extent shrinks. Multi-view acquisition systems and rotating platforms can improve coverage, but in dense greenhouse canopies, occlusion is essentially unavoidable, and registration errors and complex acquisition logistics limit their practicality. The question was whether missing regions in an incomplete single-view point cloud could be recovered reliably before any downstream measurement was attempted.

Building the training data required careful experimental design. Of the 222 fruits, 113 had relatively complete single-view point clouds and were used to construct paired incomplete-reference samples: the researchers artificially deleted portions of each fruit’s surface, using local edge deletion at ratios of 15, 30, 45, and 60 percent, plus central deletion, while retaining the original scan as the reference. Rotation, scaling, and jittering augmented these pairs into 5,085 samples, split into 4,050 for model training and 1,035 for paired validation. The remaining 109 naturally occluded fruits were reserved as an independent real-scene test set. Because these fruits had no corresponding complete references, their reconstruction could only be assessed through visualization and downstream phenotypic accuracy against manual measurements, a distinction the authors emphasize throughout.

ABox-PC itself is a generative adversarial network built around a generator and a discriminator. The generator follows an encoder-decoder architecture: the encoder extracts global shape features from the partial input and produces a coarse 256-point cloud, while the decoder progressively refines the output through two refinement stages, expanding 512 seed points selected by farthest point sampling into 1,024 and then 2,048 points. The discriminator, trained on convolutional layers, multi-head self-attention, and adaptive max pooling, learns to distinguish generated clouds from real reference scans, pushing the generator toward plausible tomato geometry. Training combined a multi-stage Chamfer distance loss, applied at the coarse, intermediate, and final stages with weights annealed during training, and an adversarial loss that stabilized the output distribution.

The most distinctive component is the lightweight spatial transformer, or LST module, which addresses a problem specific to single-view greenhouse data: uneven point density. Rather than relying on a fixed neighborhood size across the whole dataset, the module estimates a local spatial scale for each query point from the average distance to its neighbors, then constructs an axis-aligned box that expands in sparse regions and contracts in dense ones. Neighboring points inside the box are grouped, their relative positions encoded, and their features aggregated through attention. Ablation experiments showed the value of this design: replacing the adaptive box with a fixed box size of 0.1 increased the average L1 Chamfer distance by 14.62 percent, with the penalty growing as deletion severity increased, while positional encoding and an eight-neighbor grouping each proved important for consistent performance.

The quantitative results are striking. On the paired-validation subset, ABox-PC achieved a squared L2 Chamfer distance of 784.76 square millimeters, the lowest among all tested models, reducing error by 57.72 percent relative to the strong CRA-PCN baseline and by more than 90 percent relative to the PointAttN baseline. Component ablations showed that self-normalization alone cut baseline error by 30.81 percent, while the LST module and adversarial learning individually reduced it by roughly 86 and 82 percent, with the full combination performing best. On the generic PCN benchmark of eight object categories, ABox-PC achieved an average L1 Chamfer distance of 6.53, comparable to AdaPoinTr and competitive with leading methods, indicating the architecture transfers beyond tomatoes without claiming state-of-the-art status on thin, branching objects such as lamps and tables.

Crucially, the improvements carried through to actual phenotyping. For the 109 naturally occluded fruits, the coefficient of determination for horizontal diameter estimation rose from 0.47 before completion to 0.79 after, and for vertical diameter from 0.57 to 0.78, with root mean square errors falling to 5.32 and 5.07 millimeters respectively. A paired statistical test confirmed a significant reduction in vertical diameter error, with a Holm-adjusted p-value below 0.001. The team also introduced an axis-weighted oriented bounding box method that downweights the sparse depth axis during principal-direction estimation, which outperformed both standard axis-aligned and oriented bounding boxes across all conditions.

Weight estimation showed perhaps the most dramatic gain. Using a spheroid volume formula combined with an empirical fruit density of 0.93 grams per cubic centimeter, drawn from historical breeding records independent of the study fruits, the direct physical estimate improved from an R-squared of 0.53 to 0.83 after completion, with root mean square error dropping from 48.76 to 29.77 grams and mean absolute percentage error from 20.12 to 10.26 percent. Because the formula and density value were held constant, the gains stem purely from better geometric inputs. Sensitivity analysis across densities from 0.90 to 0.95 grams per cubic centimeter showed the results were robust within that narrow, cultivar-specific range, though the authors caution that applications beyond Jinpeng101 would require recalibration.

The study also tested robustness under controlled deletion severity, comparing ABox-PC against spherical fitting, ellipsoidal fitting, mirror-symmetry completion, and CRA-PCN. The geometry-assumption methods degraded sharply at 45 and 60 percent deletion, while ABox-PC maintained the best horizontal-diameter accuracy at all four deletion levels and the best vertical-diameter accuracy from 30 to 60 percent, with weight-estimation errors propagating far less severely. The authors are candid about limitations: validation covered one cultivar, one greenhouse, one maturity stage, and one sensing configuration; the completed clouds recover only the visible surface, not the unobserved back side; and localized deviations persist under severe occlusion or when time-of-flight artifacts leave irregular points near boundaries. Processing takes roughly half a second per fruit on a consumer-grade RTX 4070 GPU, suggesting the pipeline could eventually run on handheld or robotic platforms. Future work will expand the dataset across cultivars, seasons, and devices, pair natural occlusions with geometric references from multi-view reconstruction, and explore uncertainty-aware completion, steps toward making single-shot, AI-corrected fruit measurement a routine tool in the greenhouse.

Subject of Research: Deep learning-based point cloud completion for single-view tomato fruit phenotyping in greenhouses

Article Title: ABox-PC: Adaptive box-guided completion of single-view tomato fruit point clouds for improved greenhouse phenotyping

Article References: Zhu, T., Wang, F., Feng, C., Huang, L., Mou, Q., Lu, X., Min, W., Hu, X., & Nie, Y. (2026). ABox-PC: Adaptive box-guided completion of single-view tomato fruit point clouds for improved greenhouse phenotyping. Smart Agricultural Technology, 15, Article 102625. https://doi.org/10.1016/j.atech.2026.102625

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102625

Keywords: point cloud completion, tomato phenotyping, greenhouse sensing, generative adversarial network, 3D reconstruction, RGB-D camera, deep learning, fruit diameter estimation, weight estimation, occlusion, spatial transformer, precision agriculture

Neural radiance fields, or NeRF, represent a scene as a continuous volumetric function: a neural network takes a 3D position and viewing direction and predicts a density and a view-dependent color at each point. Rendering proceeds by casting rays through the scene, sampling points along each ray, and accumulating their predicted colors weighted by accumulated transmittance, which naturally models occlusion and soft appearance transitions. Because the representation is implicit, NeRF can capture fine texture detail and complex lighting behavior without explicit mesh extraction, a property that makes it attractive for artworks whose carved patterns, material transitions, and light-and-shadow interplay are central to their meaning. Yet the authors emphasize that applying NeRF to outdoor public art is not straightforward, since uncontrolled lighting, partial occlusion, and uneven viewpoint distribution can inject contradictory supervision into training.

The core scientific contribution of the study is deliberately scoped: not a new radiance-field backbone, but a visibility-weighted cross-view supervision strategy. During preprocessing, background regions, moving occluders, saturated highlights, and strongly reflective areas are annotated as low-reliability zones and converted into binary masks. These masks are then propagated into a cross-view consistency loss: for corresponding pixels in two views, a reliability weight is computed as the product of the two views’ mask values, evaluated after projecting sampled points between cameras using estimated poses and ray depths. Only view pairs whose overlapping valid region exceeds 30 percent are retained, and correspondences must fall below a 2-pixel reprojection error. The weighted consistency term, combined with a local gradient agreement constraint, suppresses contradictory evidence from unreliable regions while the standard photometric rendering loss continues to supervise all valid pixels, scaled by a regularization coefficient of 0.05.

The experimental corpus comprised five purposively selected artworks spanning diverse materials and environments: an outdoor bronze sculpture, an indoor acrylic-and-metal installation, a stone relief, a stainless-steel and wood installation, and a mixed-material work. Each was captured from 65 to 120 viewpoints using three acquisition modes: handheld circular shooting for medium outdoor pieces, tripod-and-slider rigs for indoor installations and reliefs, and drone-based multi-orbit flight for large plaza works. Canon EOS cameras with high dynamic range capability handled ground photography, while DJI platforms covered aerial orbits, with all images captured at resolutions of at least 3840 by 2160 pixels. Roughly 80 percent of images trained the model and 20 percent were held out as unseen test views, separated by viewpoint order rather than random frame selection so that the test set contained both frontal and oblique observations. Crucially, every compared method used identical images, camera poses, and test views.

Preprocessing involved conventional but essential stabilization steps. Frames with severe motion blur, dominant occlusion, or unrecoverable overexposure were excluded, and channel-wise z-score and min-max normalization reduced scale differences across cameras and lighting conditions. The authors are careful to attribute no novelty to these operations; the methodological weight rests on the reliability masks and their propagation into cross-view weighting. The network itself used a dual-branch architecture separating density estimation from color regression, hierarchical coarse-to-fine sampling, and Fourier positional encoding to capture high-frequency geometric and texture detail. Training used a batch size of 2048 rays, an initial learning rate of 5 times 10 to the minus 4, the Adam optimizer, and roughly 200,000 iterations, with the color branch frozen early in training so that stable spatial structure could be learned first.

The quantitative results are striking within the study’s protocol. The proposed configuration achieved a normalized geometric error of 1.20 plus or minus 0.20 percent, compared with 3.80 plus or minus 0.38 percent for a mesh-based structure-from-motion and multi-view stereo pipeline and 2.40 plus or minus 0.26 percent for depth-map synthesis. On held-out views, the method reached a peak signal-to-noise ratio of 28.46 plus or minus 0.70 dB and a structural similarity index of 0.921 plus or minus 0.010, against 23.08 dB and 0.809 for the mesh pipeline and 25.14 dB and 0.867 for depth-map synthesis. Multi-view Recall, measuring the proportion of regions consistently recovered across eight held-out views per artwork, reached 96.4 percent versus 84.7 and 78.2 percent for the classical baselines. Exploratory paired t-tests across the five matched cases yielded p-values below 0.001 for all comparisons, though the authors stress these describe consistency within the selected cases rather than population-level superiority.

Qualitative inspection reinforced the numbers. On the stainless-steel and wood installation captured under direct afternoon sunlight, the standard NeRF baseline showed visible color diffusion and loss of edge definition at material boundaries, while the reliability-aware configuration preserved clearer separation between adjacent surfaces and produced fewer view-dependent artifacts in partially occluded regions. A 12-participant visual-expressiveness panel rated the NeRF outputs at a mean of 4.58 on a five-point scale, with lower mean squared error and higher edge sharpness than either classical pipeline. The ablation analysis traced each component’s contribution: mask guidance lifted PSNR from 27.72 to 28.05 dB, cross-view consistency regularization raised multi-view Recall from 93.2 to 95.1 percent, and adaptive rendering increased frame rate from 21 to 27 frames per second while cutting initial loading latency from 1.8 to 1.3 seconds.

The deployment layer translated the learned radiance fields into a browser-based interactive system built on WebGL and Three.js. Reconstructed outputs were reorganized into multi-resolution assets and cached through a manifest, with display level controlled by device load, recent frame time, interaction velocity, and local texture complexity. During rapid rotation or zooming the system serves coarse tiles to maintain responsiveness, then progressively refines detail once the viewpoint stabilizes. A pseudo-augmented-reality alignment mechanism matches the user’s real-time camera viewpoint with the reconstructed model’s observation vector through frustum projection, and hotspot regions trigger artwork descriptions, material annotations, and audio commentary. A 24-participant evaluation drawn from art and design education, cultural heritage administration, and urban public-space practice produced an overall rating of 4.42 plus or minus 0.34, with operational intuitiveness scoring highest at 4.58 and immersion showing the widest dispersion at 4.28.

The study is candid about its limits, which is part of what makes it a useful template for the field. The five artworks were purposively selected, so the sample is not statistically representative; the geometric reference was quality-controlled photogrammetry rather than laser-scanned metrological ground truth; and modern neural baselines such as Instant-NGP, Mip-NeRF 360, and 3D Gaussian Splatting were not executed on the same dataset, an absence the authors report openly rather than obscure behind classical comparators. The user study lacked documented order counterbalancing, and comparator-level dispersion was unavailable for inferential testing between visualization systems. Reliability masks, the authors note, reduce contradictory supervision but cannot recover surfaces that were never observed, leaving specular metal, translucent materials, and severe backlighting as open challenges.

Even with those caveats, the work signals where digital cultural heritage is heading. Museums, university digital humanities courses, and community cultural tourism projects all depend on high-quality, low-latency 3D content, and the demonstrated combination of sub-1.5-percent geometric error, 27-fps browser rendering, and 1.3-second loading suggests that NeRF-based pipelines can move from research demos toward practical exhibition infrastructure. The authors call for future evaluations that run Instant-NGP and 3D Gaussian Splatting under identical acquisition, pose, and test-view conditions, validate absolute geometry against laser scanning, and include counterbalanced, preregistered user studies. If those confirmations arrive, the reliability-aware strategy described here could become a standard ingredient in how cities preserve, share, and reinterpret the sculptures that animate their public spaces.

Subject of Research: Neural radiance field reconstruction and interactive browser visualization of public art sculptures

Article Title: High fidelity three dimensional reconstruction and interactive visualization of public art sculptures using neural radiance fields

Article References: Shen, D. (2026). High fidelity three dimensional reconstruction and interactive visualization of public art sculptures using neural radiance fields. Discover Artificial Intelligence, 6(1), Article 1490. https://doi.org/10.1007/s44163-026-02441-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02441-3

Keywords: neural radiance fields, 3D reconstruction, public art, sculpture digitization, interactive visualization, WebGL, Three.js, photogrammetry, cultural heritage, novel view synthesis, computer vision, digital exhibition

The technique addresses a fundamental bottleneck in X-ray fluorescence imaging, one of the most widely used tools for determining which chemical elements are present in a sample and where they are located. When an X-ray beam strikes a material, the atoms within it absorb energy and then re-emit fluorescent X-rays at characteristic energies, each one acting as a fingerprint for a specific element. By detecting these emissions, scientists can identify the presence of copper, zirconium, silver, or any other element without destroying the sample. The problem arises when researchers want a full three-dimensional picture. Traditional approaches require scanning the object point by point, measuring the fluorescence at each individual location, and repeating the process across many angles to build up a volumetric reconstruction. For large or intricate samples, that process can take considerable time and generate enormous volumes of data.

The Bar-Ilan-led team took a fundamentally different approach. Instead of measuring every point individually, they placed a randomly patterned tungsten mask in front of the X-ray beam. This mask imposes a structured pattern on the illumination, so that each measurement encodes information from multiple parts of the sample simultaneously. Rather than a series of isolated point readings, the detector captures compressed, overlapping signals that contain information about the whole volume in a form that is not immediately readable but is mathematically recoverable. The researchers then rotated a sample containing copper, zirconium, and silver while measuring the fluorescent X-rays emitted by the different elements, and a reconstruction algorithm pieced together the three-dimensional elemental structures directly from the compressed data.

The results were striking. The researchers were able to eliminate more than 97 percent of the measurements that would normally be required, yet still accurately reconstruct the locations of microscopic metal wires and particles within the sample. In practical terms, that means the same chemical map that once demanded a long, data-heavy scanning session can be recovered from a small fraction of the exposure. Adi Ben-Yehuda, a PhD student in Prof. Sharon Shwartz’s laboratory at Bar-Ilan University’s Department of Physics, described the surprise the team felt at how much information could be recovered from so little data. Instead of measuring every point separately, she explained, the method uses a structured X-ray pattern and lets the reconstruction algorithm assemble the information, a different way of thinking about how an X-ray image can be created.

The underlying idea belongs to a family of techniques known as ghost imaging, in which the image-bearing information is split between two correlated signals and recovered computationally rather than recorded directly. The new work builds on the team’s previous demonstration of two-dimensional X-ray fluorescence ghost imaging, extending the concept into a full three-dimensional imaging system. In the two-dimensional version, the structured illumination allowed the researchers to recover elemental maps without ever forming a conventional image on the detector. The leap to three dimensions required combining that compressed-sensing philosophy with tomographic reconstruction, the same mathematical framework that underlies computed tomography in medicine, but applied to fluorescence data rather than simple transmission. The result is a system in which the illumination pattern, the rotation of the sample, and the reconstruction algorithm work together as a single designed pipeline.

Why does speed matter so much in this field? The answer lies largely in the economics and physics of synchrotron facilities, the enormous particle-accelerator complexes where scientists use extremely bright X-ray beams to study materials at microscopic scales. Beamtime at these facilities is scarce and highly competitive, and every hour of measurement is precious. A technique that reduces the required measurements by more than 97 percent effectively multiplies the amount of science that can be extracted from a given allocation of beamtime. Faster imaging also opens experimental possibilities that are simply impractical today: scanning larger samples that would otherwise take days to map, following chemical processes as they unfold in real time, and studying delicate biological materials while reducing their cumulative exposure to damaging X-ray doses.

The potential applications span an unusually broad range of disciplines. In materials science, three-dimensional elemental maps reveal how alloys, catalysts, and electronic components are structured at the microscopic level, information that guides the design of stronger and more efficient materials. In green energy research, understanding the chemical distribution inside batteries and fuel cells as they charge and discharge is central to improving performance and longevity. In biology, researchers could map the distribution of trace elements within cells and tissues, where metals such as zinc, iron, and copper play essential roles in enzymes and signaling. In geology, elemental tomography can expose the growth history of minerals and the microstructure of rocks that record the planet’s past. For all of these fields, a 43-fold acceleration in acquisition time changes not just efficiency but the very scope of what can be attempted.

Prof. Sharon Shwartz of Bar-Ilan University framed the advance as a change in a long-standing trade-off. X-ray imaging has traditionally forced researchers to balance the amount of information they want against the time and data required to obtain it, he noted, but the new approach shows that this trade-off can be altered. By designing the X-ray illumination together with the computational reconstruction, the team can recover detailed three-dimensional chemical information from a remarkably small number of measurements. That co-design principle, in which hardware and software are developed as an integrated system rather than as separate stages, reflects a broader trend in modern imaging science, where clever mathematics increasingly substitutes for brute-force data collection.

The study was conducted with Dr. Alexander Rack of the European Synchrotron Radiation Facility and Dr. Nicola Viganò of CEA and Université Grenoble Alpes, and was supported by the Israel Science Foundation and the European Synchrotron Radiation Facility. The combination of an academic physics laboratory with two major European photon facilities underscores the kind of collaboration this technique will need as it matures, since structured-illumination experiments depend on precisely engineered masks and stable, brilliant X-ray sources of the kind found only at synchrotrons.

The researchers are already looking ahead. Their next steps include applying the technique to more complex biological and geological samples, whose heterogeneous structures and weaker elemental signals will test the robustness of the reconstruction algorithm, and incorporating artificial intelligence and deep learning to further improve imaging speed and resolution. If those efforts succeed, the compressed-measurement philosophy demonstrated here could become a standard tool at light sources around the world, turning what was once a slow, data-hungry mapping process into something fast enough to watch chemistry happen in three dimensions, live.

Subject of Research: Compressive three-dimensional X-ray fluorescence ghost tomography for rapid chemical imaging

Article Title: New X-ray imaging method maps 3D chemical structures 43 times faster

Article References: New X-ray imaging method maps 3D chemical structures 43 times faster. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: X-ray fluorescence, ghost imaging, tomography, Bar-Ilan University, synchrotron, compressed sensing, chemical imaging, Optica, tungsten mask, 3D reconstruction, materials science, beamtime efficiency

The research, led by Huaqing Huang and Naifei Liu of the School of Design at Shanghai Jiao Tong University, together with colleagues from Tsinghua University, the University of Virginia, and industry partners, tackles a problem that has long divided the virtual heritage community. On one side stands scientific rigor: reconstructions must be grounded in verifiable evidence, from archaeological surveys to historical documents, and must clearly distinguish what is known from what is inferred. On the other side stands contextual narrative: an immersive display must feel complete and coherent, giving visitors a believable sense of place rather than a scattering of fragmentary, heavily caveated ruins. The authors describe these two dimensions as decoupled in much existing practice, and their central contribution is a workflow designed to reunify them.

That workflow unfolds in four sequential stages: data collection and processing, spatial scripting, virtual reconstruction, and asset production. Each stage addresses a distinct technical challenge. Data collection aggregates the heterogeneous evidence base, including surveying data, heritage documentation, maps, and interdisciplinary historical materials. Spatial scripting then organizes this material into a structured plan for the virtual environment, determining how spaces relate to one another and how a visitor’s journey through them will unfold. Virtual reconstruction translates the scripted spatial logic into three-dimensional form, while asset production refines the models, textures, and interactive elements that make the environment usable in a real-time VR platform.

One of the most consequential questions the team confronted was prioritization. Urban heritage on the scale of Jingdezhen’s porcelain industry cannot be reconstructed in exhaustive detail everywhere at once; computational budgets, production timelines, and the limits of historical evidence all impose constraints. The study therefore addresses how to determine which elements of the 3D recreation deserve the highest fidelity and which can be represented more schematically, in order to balance scientific validity against contextual completeness. This triage process, the authors argue, is where interdisciplinary judgment matters most: archaeologists, historians, conservation specialists, and digital designers must negotiate together what the virtual city should emphasize and where approximation is acceptable.

The case at the heart of the project is the core heritage area of Jingdezhen’s porcelain industry, the zone where kilns, workshops, and the commercial infrastructure of ceramic production were most densely concentrated. Reconstructing an industrial urban landscape poses particular difficulties compared with the monuments that dominate virtual heritage work. Industrial heritage is often fragmentary, with above-ground remains that are partial, altered, or buried, and its significance lies as much in processes and spatial relationships as in individual structures. Conveying how clay moved through a production chain, how kilns were organized within the city, and how workers and merchants inhabited these spaces requires a reconstruction strategy that encodes function, not just geometry.

The scale of the collaboration is itself notable. The project involved an interdisciplinary team of more than twenty experts and spanned over a year of work. The acknowledgements reveal the breadth of institutional support required: the Jingdezhen Ceramic Culture Tourism Group, known as Taowenlyu, provided early surveying data and helped coordinate implementation on the display platform; Tsinghua University and the Qingyuan Conservation team supplied key heritage data and academic consultation; and Zhongxin Digital Culture defined system technical requirements while advancing technical implementation and operational activities in parallel. The research itself was supported by the Spatial Intelligence and Virtual-Real Fusion Design Laboratory at Shanghai Jiao Tong University, where the academic core of the team is based.

This division of labor reflects a broader trend in digital heritage, where universities, cultural institutions, and technology companies increasingly collaborate on projects that no single organization could deliver alone. Heritage datasets of the kind held by conservation teams are rarely formatted for real-time rendering; game-engine-ready environments demand optimization decisions that can quietly alter historical content if made without expert oversight. The Jingdezhen workflow formalizes the handoffs between these communities, making explicit where historical interpretation ends and technical production begins, and building checkpoints into the pipeline where domain specialists can verify that the emerging 3D environment remains faithful to the evidence.

The integration of interdisciplinary historical materials, which the paper identifies as a core question of the process, is arguably the methodological heart of the work. Historical sources for a city like Jingdezhen arrive in incompatible forms: textual records, historic maps, archaeological excavation reports, architectural surveys, photographs, and oral accounts of craft practice. Each carries its own scale, precision, and bias. Fusing them into a single coherent spatial model requires decisions about alignment, dating, and confidence that are as much historiographical as they are technical. The study’s treatment of this integration stage suggests a model for other cities pursuing similar projects, particularly those whose heritage value lies in urban patterns and industrial systems rather than in freestanding architecture.

The timing of the project is significant as well. Jingdezhen’s pursuit of World Cultural Heritage status has made the legible presentation of its porcelain industry a matter of international importance, and immersive VR display offers a way to communicate that value to audiences who cannot physically access fragile sites or who would otherwise see only fragmentary remains. Virtual reconstruction, in this framing, is not merely a visualization exercise but an argument: it assembles scattered evidence into a spatial narrative that makes the case for the city’s outstanding universal value. The VR project described in the study was developed explicitly as part of that bid, positioning digital heritage tools within the formal machinery of UNESCO nomination processes.

The authors are candid that the approach carries challenges alongside its potentials. Virtual recreation at urban scale remains expensive, slow, and dependent on the availability and quality of underlying data. Interpretive choices made during spatial scripting and reconstruction can harden into apparent fact once rendered in convincing three dimensions, a risk the field of virtual heritage has debated for decades. By publishing the workflow, the team aims to make those choices visible and repeatable, so that other projects can adopt, critique, and improve the method. The article, published open access in npj Heritage Science on 11 October 2026, arrived as a peer-reviewed accepted manuscript carrying a permanent DOI, and its authors declare no competing interests. As cities worldwide grapple with how to document industrial and urban heritage threatened by development, decay, and climate pressures, the Jingdezhen project offers a template: a disciplined pipeline that treats historical evidence, narrative design, and real-time technology as equal partners in the work of keeping a vanished city alive in memory and in silicon.

Subject of Research: Virtual reality reconstruction of the urban porcelain-industry heritage of Jingdezhen, China

Article Title: Virtual recreation of urban heritage of Jingdezhen for immersive VR display

Article References: Huang, H., Liu, N., Zhang, K., Wang, S., Zhu, S., Yue, K., & Li, Z. (2026). Virtual recreation of urban heritage of Jingdezhen for immersive VR display. npj Heritage Science. https://doi.org/10.1038/s40494-026-03036-y

Image Credits: AI Generated

DOI: 10.1038/s40494-026-03036-y

Keywords: virtual reality, digital heritage, Jingdezhen, porcelain industry, urban heritage, 3D reconstruction, World Cultural Heritage, npj Heritage Science, spatial scripting, immersive display, cultural heritage, Shanghai Jiao Tong University

The scale of the problem is easy to underestimate because the failure mode is nearly invisible. When a pinhole calibration is fitted to data collected through a refractive interface, the optimization does not crash or warn the user; instead it absorbs much of the mismatch into an inflated focal length and spurious lens distortion. Earlier studies by Sedlazeck and Koch, and by Jordt and colleagues, characterized this behavior for cameras mounted close behind underwater ports, finding that residual triangulation error grows with the camera’s standoff from the interface and with the distance to the imaged point, eventually reaching several percent of the camera-to-point distance. Jordt’s group observed depth errors of roughly 20 centimeters for surfaces at 2 meters. Those studies examined standoffs of centimeters; a fixed camera array imaging an entire aquarium tank from a meter or more away operates an order of magnitude beyond that range, where the errors grow in direct proportion to the standoff.

What makes the bias especially dangerous is that the checks practitioners usually run cannot detect it. Distances between nearby reconstructed points remain sub-millimeter under both a pinhole model and a corrected one, across a wide range of detection noise. Revealing the systematic depth-axis error requires an absolute-position check or a held-out rigid object spanning a long depth baseline, measurements that few pipelines routinely make. The affected fields are broad, spanning behavioral biology, aquaculture monitoring, underwater archaeology, and robotic navigation, and once the bias enters at calibration it silently contaminates everything reconstructed from the model.

Refraction-aware calibration theory is itself well established. Treibitz and colleagues showed that flat-interface refraction produces depth-dependent distortion that no radial polynomial can capture across the imaging volume, Agrawal and collaborators derived the general multi-layer flat-refraction theory, and Chadebecq’s team extended refractive geometry to two-view reconstruction. Recent systems such as Refractive COLMAP, the CalibMar toolbox, and a virtual-imaging method for binocular laser scanners build on these foundations. But all three estimate a separate interface distance for each camera, treating one physical plane as many. For a fixed array sharing a single water surface, that per-camera approach leaves each camera’s position along the interface normal redundant with its own interface height, an estimate that Treibitz’s work shows to be ill-conditioned unless the target reaches the edge of the frame, where refraction is strongest. No existing open-source tool was built to exploit the shared interface as geometric structure.

AquaCal, developed by Tucker Lancaster and Patrick McGrath at the Georgia Institute of Technology, fills that gap. It is a pip-installable Python library that models refraction at a single planar interface explicitly, applying Snell’s law in both projection directions, and folds the physics into one joint bundle adjustment over the camera extrinsics, a single global interface height, and the per-frame placement of the calibration target. The world frame is oriented with its Z axis along the interface normal, so the interface becomes the single horizontal plane at a height the software calls water_z. Each camera’s physical standoff, the gap that enters Snell’s law, is then derived directly from that one shared parameter. In synthetic tests, replacing the shared parameter with a per-camera estimate placed one physical surface at twelve different heights spanning 11 millimeters, and recovered every camera’s standoff less accurately across all ten random seeds tested.

The calibration runs in three stages, each designed around the conditions under which its parameters are best estimated. Stage 1 estimates lens intrinsics from in-air video of a small ChArUco board, a chessboard pattern embedded with ArUco fiducial markers, using OpenCV and supporting standard, rational, and fisheye lens models. This separation matters because under refraction a camera’s focal length and its standoff trade off against each other, making a joint fit from underwater observations alone poorly conditioned, whereas in air the projection is unambiguous. Stage 2 seeds the final joint solve, which is non-convex and converges reliably only from a starting point already close to the solution, by chaining pairwise relative camera positions and fusing the redundant links through rotation averaging, without requiring any single board placement visible to every camera at once. Stage 3 then estimates everything together, minimizing the refractive reprojection error across all observations, with an optional second pass that unlocks each camera’s focal length and principal point only after the geometry has converged.

The engineering beneath the pipeline is as carefully considered as the geometry. AquaCal computes the Jacobian of its optimization by finite differences over an explicit block-sparse pattern, applying Curtis–Powell–Reid column grouping so that the number of evaluations is a property of a single observation rather than of the array, meaning the savings grow as cameras and, more steeply, frames are added. A Huber robust loss down-weights outlier corner detections without a separate filtering stage. Cameras with high-distortion optics or limited target visibility can be designated auxiliary, registered after the fact against the frozen shared parameters so they cannot perturb the primary solution, and flagged in the output so downstream users can weigh their accuracy cost. The whole optimization scales to rigs of a dozen or more cameras in under twenty minutes on a workstation, with peak memory use near 11 gigabytes, and terminates at a cost tolerance of 1e-8 rather than an iteration cap.

The validation results are striking. In synthetic tests of an idealized twelve-camera rig about a meter above a tank, with 0.5 pixels of Gaussian noise added, the refractive model held focal lengths near ground truth with a mean absolute drift of just 0.066 percent, while the non-refractive baseline drifted by 6.83 percent, consistently overestimating focal length to compensate for the apparent-depth compression that refraction introduces. The refractive model drove reprojection error to the noise floor at 0.498 pixels, while the baseline plateaued about 2.5 times higher. The gap widened dramatically in 3D reconstruction. Triangulating held-out board placements at depths from 0.07 to 1.47 meters below the surface, the refractive model’s depth-axis error stayed flat between 1.7 and 1.9 millimeters across the entire tested volume, while the baseline climbed from 17.6 millimeters to 243 millimeters at the farthest point, roughly 2.5 meters from the reference camera, an approximately 132-fold degradation. Local scale accuracy told a complementary story: the refractive model held inter-corner distance error flat between 0.38 and 0.42 millimeters at all depths, while the baseline traced a U-shaped curve, accurate only within the band the calibration target had swept.

On the physical rig, a thirteen-camera array mounted on a hexagonal ceiling structure about a meter above a 2460-liter tank with an open water surface, the primary cameras achieved sub-pixel mean reprojection accuracy of 0.82 pixels, and held-out triangulated inter-corner distances showed a mean absolute error of 0.258 millimeters, a 0.43 percent mean relative error across 7,762 comparisons, with minimal systematic scale bias. The rig is not a test bed but a production platform for continuous behavioral monitoring of freely swimming Lake Malawi cichlids, supporting multi-fish 3D positioning, inter-individual distance measurements, and animal–substrate interaction analyses. Two downstream libraries, AquaMVS for refraction-aware multi-view stereo of the tank substrate and AquaPose for 3D pose estimation of swimming fish, already consume AquaCal camera models directly, and manuscripts describing them are in preparation.

The implications reach well beyond aquaria. Because the model parameterizes refraction by an arbitrary ratio of refractive indices, the pipeline extends without modification to non-aqueous media such as oil baths, with reconstruction accuracy holding across indices from 1.333 to 1.55, and a side-mounted rig imaging through a vertical interface requires no change to the model. The software, released under an MIT license with full documentation, archived input data and results on Zenodo, and Jupyter tutorials introducing both the pipeline and the refraction problem itself, replaces the improvised one-off corrections laboratories have historically relied on with a standardized, comparable implementation. The authors point to multi-layer and curved interfaces, such as dome housings, and online recalibration from tracked-animal correspondences as the most immediate extensions, both of which sit within the existing optimization architecture. For any laboratory counting on cameras to measure the underwater world accurately, the message is clear: the bending of light at the water’s surface is no longer an acceptable unknown.

Subject of Research: Refraction-aware multi-camera calibration for imaging through planar air–water interfaces

Article Title: AquaCal: refraction-aware multi-camera calibration for planar air–water interfaces

Article References: Lancaster, T., & McGrath, P. (2026). AquaCal: refraction-aware multi-camera calibration for planar air–water interfaces. SoftwareX, 36, Article 103091. https://doi.org/10.1016/j.softx.2026.103091

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103091

Keywords: camera calibration, refraction, Snell's law, 3D reconstruction, bundle adjustment, multi-camera arrays, underwater imaging, computer vision, open-source software, behavioral biology, aquaculture, ChArUco board

Phenotyping, the quantitative measurement of plant traits such as height, stem thickness and architecture, is the backbone of modern crop breeding. Breeders rely on it to predict yields, assess stress tolerance and select improved varieties, but traditional manual measurement is slow, labour-intensive and often destructive. Indoor scanning chambers can achieve micron-level precision by rotating plants past banks of sensors, yet they can only handle small batches under controlled lighting. Field conditions are far messier: sunlight shifts minute by minute, leaves overlap into dense canopies, and the repetitive rows of nearly identical plants confuse algorithms that depend on distinctive visual features.

Existing field platforms each carry trade-offs. Drones cover large areas fast but see only the top of the canopy, missing the stems and lower stalks that determine whether a plant will resist lodging, the stem breakage that devastates harvests. Rail-mounted scanners deliver repeatable measurements but demand costly permanent infrastructure. Ground vehicles can drive between rows and peer under the canopy, yet most published systems were validated only on young seedlings and depend on offline processing pipelines such as structure-from-motion or iterative closest point registration, which can take tens of hours to reconstruct a single field plot.

The new platform, developed by researchers including Si Yang and Xinyu Guo, attacks these weaknesses directly. A high-clearance vehicle with a 145-centimetre wheelbase straddles two rows of maize at a time, carrying three Orbbec Femto Bolt RGBD cameras: one looking straight down at the canopy and two angled at roughly 60 degrees to capture the sides of the plants. The cameras are hardware-synchronised through a dedicated sync hub to within 100 microseconds, which at the vehicle’s 0.2-metre-per-second cruising speed corresponds to a displacement of just 0.02 millimetres between frames. Crucially, the system needs no GPS or inertial measurement unit, relying entirely on vision.

The heart of the innovation is the way the software treats the three cameras as a single rigid sensing unit. Before deployment, a checkerboard calibration fixes the exact rotation and translation between the cameras, and these relationships are locked into the tracking and mapping modules of the SLAM system. Instead of each camera navigating alone, the framework merges features from all three viewpoints into one unified set, so if one camera stares at a featureless leaf surface, the others can still anchor the robot’s position. Keyframe insertion is also forced whenever any single camera drops below five matched features, a safeguard that preserves tracking when one view is temporarily blinded by glare or occlusion.

Feature extraction itself was redesigned for the peculiar hostility of a cornfield. Standard ORB detectors tend to cluster points on high-contrast soil or sky, leaving the weakly textured leaves starved of landmarks. The team’s solution uses the Excess Green index, a simple colour formula that separates vegetation from background, to divide each image into crop and non-crop zones. Fine grids concentrate feature detection on the plants, coarse grids cover the rest, and a semantic weighting scheme boosts the priority of crop features across the image pyramid. An adaptive FAST threshold, computed from local grey-level statistics in each grid cell, allows corners to be detected even in flat, unevenly lit foliage.

The results are striking. In ablation tests on the hardest datasets, mature V10 to V12 stage maize under strong sunlight and wind, a conventional single-camera ORB-SLAM3 baseline managed tracking continuity as low as 8 percent, and failed outright in six of ten test sequences. The complete framework achieved 100 percent tracking continuity on every sequence, regardless of illumination, weed pressure or plant motion. Where both systems worked, the new method kept roll-angle drift to around 1.7 degrees on average, compared with swings of more than 33 degrees for the baseline, and maintained height stability of roughly 0.18 metres over runs of up to 60 metres.

Speed matters as much as robustness. Reconstructing 1,300 frames with a structure-from-motion pipeline took about 36 hours, and a SIFT-plus-ICP approach around 26 hours, both offline. The new system maps in real time on a CPU alone as the robot drives. Against reference scans of 30 potted maize plants, its reconstructions showed a mean Chamfer distance error of about 2.18 centimetres, accurate enough for field trait extraction. Downstream, an automated pipeline segments individual plants by fusing bird’s-eye density projections with 3D clustering, then measures height and models each stem cross-section as an ellipse rather than a circle, reflecting the true geometry of a maize stalk.

Validation against hand measurements of 259 field-grown plants showed the system estimating plant height with a coefficient of determination of 0.85 and a root mean square error of about 41 millimetres. Stem diameter estimates along the major and minor axes reached R-squared values of 0.79 and 0.76 with errors of roughly 1.2 and 1.0 millimetres, comparable to or better than previous methods. Across ten field strips totalling nearly 3,900 plants, automated counting matched manual counts with a mean accuracy of 94.55 percent, all from data collected across four growth stages, sunny and cloudy skies, and fields with and without chemical weed control.

The authors are candid about limits. Cumulative drift over long distances will eventually require loop closure or RTK-GNSS fusion, the fixed camera calibration may loosen under vibration and temperature swings, and the simple colour-based vegetation mask cannot distinguish maize from large green weeds. Future versions will adopt lightweight deep-learning segmentation and visual-inertial navigation. Even so, the demonstration marks a turning point: a robot that can drive down a row of mature maize and hand breeders a complete, quantified 3D census of every plant before it reaches the end of the field. For breeding programmes racing to develop higher-yielding, lodging-resistant crops, that could compress seasons of laborious measurement into a single afternoon pass.

Subject of Research: A multi-RGBD SLAM framework on an unmanned ground vehicle for real-time in-situ 3D phenotyping of field maize

Article Title: UGV based multi-RGBD SLAM framework for high-throughput in-situ 3D phenotyping of field maize

Article References: Yang, S., Liang, Y., Huang, G., Qiu, G., Wen, W., Wang, C., Gou, W., Guo, X., & Zhao, C. (2026). UGV based multi-RGBD SLAM framework for high-throughput in-situ 3D phenotyping of field maize. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.08.019

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.08.019

Keywords: maize phenotyping, SLAM, RGBD cameras, unmanned ground vehicle, 3D reconstruction, precision agriculture, plant breeding, stem diameter, plant height, semantic feature extraction, high-throughput phenotyping, agricultural robotics

The difficulty of seedling phenotyping is easy to underestimate. Lettuce seedlings are diminutive, morphologically plastic, and in some cultivars spectrally almost identical to their growing medium. In ordinary RGB images, this produces low visual contrast and blurred boundaries between plant and substrate, undermining the appearance-based algorithms that work well on mature crops. Yet the seedling stage is precisely when subtle morphological changes carry the most biologically significant information about genotype, vigor, and stress response. Single-modality sensing compounds the problem: color cameras miss physiological state, while depth sensors alone struggle with the weak textures of young foliage. What was needed was a system that could capture geometric structure and physiological condition simultaneously, at scale, without touching a single leaf.

The team’s answer is an unmanned ground vehicle with a straddle-type chassis that drives directly over rows of potted lettuce, imaging the canopy from above without disturbing growth. Mounted at the center of the vehicle is a rigidly coupled, synchronously triggered imaging unit combining a time-of-flight Femto Bolt depth camera with a CropEye-A1 multispectral camera that records reflectance in four narrow bands centered at 550, 660, 720, and 840 nanometers. These bands are critical proxies for chlorophyll content, biomass, and plant stress. The UGV captures one synchronized frame every four seconds as it follows predefined paths, a deliberately sparse sampling strategy that balances spatial overlap against storage and computational cost, with mechanical dampers suppressing the blur of vehicle motion.

Because the two sensors differ in resolution, field of view, and optical geometry, the raw data streams arrive misaligned. The researchers built a registration module that extracts SIFT keypoints from both modalities, generates initial matches through FLANN descriptor pairing, and then filters false correspondences using a statistical slope- and angle-consistency test based on the median absolute deviation criterion. Surviving matches feed a RANSAC estimator that computes an affine transformation, registering each multispectral band precisely to the RGB reference frame. This pixel-level alignment is the quiet foundation of the entire pipeline: without geometrically consistent RGBD and multispectral data, every downstream step would inherit systematic error.

At the heart of the framework sits MS-SegNet, a novel dual-backbone instance segmentation network that extends the lightweight YOLOv11n-Seg architecture. One backbone processes RGBD input while a parallel branch handles multispectral data, allowing each to specialize in its own modality before features are merged. Fusion is hierarchical rather than uniform: a Shallow Cross-Self Attention Fusion module combines edge-level features at the C3 and C4 stages, using cross-attention, CBAM-style self-attention refinement, and a learnable prompt vector that adaptively weighs geometry-dominant against reflectance-dominant information. At the deepest stage, a Transformer-based Deep Fusion Module applies local-window attention to align high-level semantic representations. The backbone itself was upgraded with an ADown downsampling module that preserves fine leaf detail, and a C2PSDA attention block that contrasts two attention heads to suppress the false responses that substrate textures otherwise provoke.

The numbers are striking. On a dataset of 410 synchronized multimodal frames covering 237 lettuce varieties across seven horticultural types, with 1,726 manually annotated seedling instances, MS-SegNet achieved a precision of 0.993, an mAP@50 of 0.994, an mAP@75 of 0.973, and an mAP@50:95 of 0.854, outperforming YOLOv8n-Seg, YOLOv9n-Seg, Mask R-CNN, and Mask2Former under identical training conditions. The two-stage generalist models lagged by 6.7 to 8.7 percentage points on the strictest metric, underscoring how demanding fine-grained seedling segmentation is. Ablation experiments confirmed that each proposed module contributes measurable gains, and Grad-CAM visualizations showed the network focusing tightly on seedling contours where baseline models scattered attention across pots, soil, and irrigation pipes. Mid-level feature fusion proved decisively better than early input stacking or late decision merging, which cannot refine boundaries through cross-modal interaction.

Reconstructing the scene in 3D posed a different challenge. With frames captured only every four seconds, inter-frame overlap is minimal and traditional visual odometry or SLAM approaches fail. The team instead initialized global camera poses with a structure-from-motion pipeline, masked out static vehicle components that would generate false feature matches, and then recovered absolute metric scale by comparing SfM-derived displacements with depth-based ICP alignments across frame pairs, aggregating robustly via the median. A final incremental generalized ICP refinement, run in a multi-resolution voxel framework with surface-normal constraints, suppressed residual drift. Qualitative comparisons showed that classical SfM-MVS produced hole-ridden, fragmented canopies, while modern monocular reconstruction models such as VGGT and Pi3 suffered severe drift and geometric distortion on the weakly textured, top-down greenhouse imagery. The proposed pipeline alone yielded metrically consistent, structurally stable reconstructions across entire trajectories.

Notably, the reconstruction framework generalized without any parameter adjustment to two morphologically contrasting species, lily and tomato, preserving tomato’s layered branching despite heavy self-occlusion and recovering lily’s smooth, repetitive stems where photometric features are inherently scarce. For the lettuce seedlings themselves, the fused semantic multispectral point clouds enabled extraction of canopy area, plant height, crown width, volume, and compactness, alongside vegetation indices including NDVI, NDRE, and PRI. System-derived plant height and crown width agreed strongly with manual measurements, achieving R-squared values of 0.8379 and 0.918 with root-mean-square errors of roughly 11 millimeters, centimeter-level accuracy sufficient for quantitative growth monitoring.

The spectral dimension adds a capability that geometry alone cannot deliver: physiological assessment. Using the 90th percentile of NDVI and NDRE distributions within each segmented seedling, a statistic that suppresses background interference and highlights the most active leaf tissue, the researchers established a four-level health grading system from vigorous to poor. Correlation analysis revealed strong positive associations between vegetation indices and volumetric growth traits, while RGB-derived intensities correlated weakly, confirming that trichromatic color carries little discriminatory power in substrate-dominated greenhouse scenes. The upshot is early detection of subtle stress signatures in sparse seedling canopies, something neither RGB imaging nor 3D geometry could achieve independently.

The authors are candid about limitations. Segmentation accuracy dips under the strictest overlap thresholds along fine leaf edges, reconstruction quality is sensitive to deviations from planned trajectories under the fixed sampling interval, and the absence of laser-scanner ground truth means geometric fidelity was validated indirectly through trait agreement. As plants mature and canopies overlap, occlusion will inevitably degrade reconstruction, and transferring the segmentation model to new species still requires fresh annotation. Even so, the framework offers a compelling template for automated, non-destructive seedling phenotyping in controlled environments. The team’s stated next steps, real-time deployment on autonomous robots, integration with temporal growth models, and expansion to additional spectral modalities and crop varieties, point toward a future in which every seedling in a breeding program carries its own continuously updated digital twin, accelerating data-driven breeding and precision cultivation at industrial scale.

Subject of Research: UGV-based multimodal RGBD–multispectral fusion for 3D phenotyping of greenhouse lettuce seedlings

Article Title: UGV-based multimodal RGBD–multispectral fusion framework enables high-quality 3D phenotyping of greenhouse lettuce seedlings

Article References: Yang, S., Qiu, G., Xia, J., Zhao, Y., Wen, W., Wang, C., Gou, W., Guo, X., & Zhao, C. (2026). UGV-based multimodal RGBD–multispectral fusion framework enables high-quality 3D phenotyping of greenhouse lettuce seedlings. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.08.015

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.08.015

Keywords: lettuce seedlings, plant phenotyping, multispectral imaging, RGBD sensing, unmanned ground vehicle, 3D reconstruction, instance segmentation, MS-SegNet, controlled environment agriculture, NDVI, greenhouse robotics, deep learning