The study, published in Global Surgical Education, the journal of the Association for Surgical Education, enrolled 48 medical students who had completed a course in operative technique and experimental surgery. Every participant was a genuine laparoscopic novice: none had undertaken formal simulation training, practiced on the study platform, or received structured instruction in laparoscopic suturing before the experiment. This naive cohort was a deliberate choice. By testing learners before they could compensate for lost depth cues with experience, the investigators could isolate the pure effect of visualization modality on early skill acquisition, the phase of training where the cognitive burden of learning instrument handling, camera interpretation, and task sequencing is heaviest.

The trial used a randomized cross-over design to control for the wide individual variability that plagues surgical skills research. Participants were allocated one-to-one to two sequences. One group performed four standardized laparoscopic tasks first under two-dimensional high-definition visualization and then repeated them under three-dimensional high-definition vision; the other group followed the reverse order. Because the intervention was a visual display, blinding of participants and investigators was impossible, but the cross-over structure allowed each student to serve as their own control. Task completion time in seconds and the number of predefined errors served as the primary objective outcomes, while a study-specific post-task questionnaire captured perceived depth perception, performance, speed, visual adaptation, and adverse symptoms.

The four tasks were calibrated to test progressively demanding psychomotor skills on a physical box trainer using real graspers and needle holders. Task one required positioning ten metal cylinders at predefined locations on a board, with an error counted for any misplaced cylinder. Task two involved transferring ten pearls, one at a time, from a container with a 42-millimeter opening into one with a tighter 35-millimeter opening. Task three demanded that participants grasp a disposable needle and its cap from the bottom of the box and cap the needle above the floor, with punctures and dropped objects counted as errors. Task four, the most complex, tested suturing and knot tying on a sponge model with 3-0 nylon on a CTI 3/8 circle needle, timed from the first stitch through the third knot.

The results were striking, particularly among students who moved from two-dimensional to three-dimensional visualization. In this sequence, three-dimensional vision produced statistically significant improvements in every single task. Cylinder positioning was completed 32.9 percent faster with 40.8 percent fewer errors. Pearl transfer was 24.6 percent faster with errors reduced by 67.6 percent. Needle capping improved by 38.5 percent in time and 32.6 percent in errors, while knots and suturing were 22.3 percent faster with 29.4 percent fewer errors. All comparisons reached statistical significance, with p values at or below 0.002 for the time outcomes and p less than 0.001 for error counts.

The reverse sequence, in which students began with three-dimensional vision and switched to two-dimensional, told a subtler story, as expected when learning from task repetition can mask the benefit of a second modality. Here, three-dimensional visualization still produced 50 percent fewer errors in pearl transfer and a 27.4 percent faster completion of needle capping with 27 percent fewer errors, both statistically significant, though some differences narrowed or lost significance. To address carryover learning directly, the investigators compared initial performance between the two groups, contrasting students who had never seen the tasks with those starting on three-dimensional vision. This first-exposure comparison favored 3D in three of the four tasks, including a 21.7 percent faster cylinder positioning with 38.8 percent fewer errors and a dramatic 50.4 percent faster needle capping with 36.9 percent fewer errors, strengthening the case that stereopsis confers a genuine advantage rather than a mere order effect.

The subjective data reinforced the objective measurements with remarkable consistency. An overwhelming 97.9 percent of participants reported better depth perception with three-dimensional vision, 79.2 percent reported better performance and more accurate gestures, and 85.4 percent reported faster task execution. A slimmer majority, 52.1 percent, found the three-dimensional display easier to adapt to, and 47.9 percent actually considered the two-dimensional image more natural, a finding the authors attribute to limited prior exposure to stereoscopic displays, variation in individual tolerance to 3D imaging, monitor positioning, and discomfort related to the glasses. Adverse symptoms were modest but real: 33.3 percent reported discomfort or hand pain during instrument manipulation, 25 percent reported blurred vision, and 8.3 percent reported dizziness, while one-third of participants experienced no symptoms at all.

The technical significance of the study lies less in the individual percentages than in the platform itself. The authors intentionally excluded virtual reality, augmented reality, and force-feedback haptics so that the experiment would isolate the effect of stereoscopic visualization while preserving the tactile interaction of real instruments within a physical trainer. The configuration combined a Sony HDR-TD20 camera capable of both 2D and 3D high-definition recording, an HDMI cable, a Samsung LED TD950 2D/3D monitor, and a standard laparoscopic box trainer with conventional graspers and needle holders. Commercial 3D laparoscopic towers and robotic platforms remain prohibitively expensive for many training programs, and the authors situate their work within a broader movement toward affordable simulation, citing recent examples including biodegradable 3D-printed simulators, validated low-cost trainers, and automated error-detection systems that improve trainee confidence.

The findings also engage with ongoing debates about validity in surgical education. The authors are careful to note that their simulator should not be described as valid in isolation; rather, the completion times, error counts, and preference responses constitute context-specific evidence about immediate simulator performance among novices. They also acknowledge a caution raised in prior literature that two-dimensional versus three-dimensional comparisons vary widely in task type, participant experience, endpoints, and subjective assessment. Among the study’s stated limitations are its exclusive novice population, the short interval between modalities that may have permitted carryover learning, the absence of delayed retention testing or transfer to clinical performance, and a study-specific questionnaire that was not formally psychometrically validated. Baseline measures of video-game use and other visuospatial experience were also not collected.

Nevertheless, the message for global surgical education is clear and potentially transformative. Where access to robotic simulators, high-fidelity virtual reality trainers, or commercial 3D towers is limited by capital and maintenance costs, a low-cost stereoscopic trainer may allow students and residents to acquire fundamental visuospatial and psychomotor skills before entering the operating room. The authors position such platforms not as replacements for emerging autostereoscopic and augmented-reality systems, which may eventually combine depth perception with image-guided overlays and remote instruction, but as a pragmatic bridge until those technologies become widely accessible. Future studies, they suggest, should include delayed retention testing, transfer tasks, clinical outcome measures, and baseline visuospatial profiling. If replicated, the demonstration that a few hundred dollars of consumer hardware can outperform flat high-definition vision on every measure that matters in early laparoscopic training could help democratize surgical skills education in precisely the settings where trained surgeons are scarcest.

Subject of Research: Three-dimensional versus two-dimensional visualization in low-cost laparoscopic skills training for novices

Article Title: 3-Dimensional versus 2-dimensional visualization on laparoscopic skills training in a low-cost simulation model: a randomized cross-over trial

Article References: Heldwein, F. L., Freire, E. S., Vicente, M. V. S., Porath, H., & Veiga, C. B. (2026). 3-Dimensional versus 2-dimensional visualization on laparoscopic skills training in a low-cost simulation model: a randomized cross-over trial. Global Surgical Education – Journal of the Association for Surgical Education, 5(1), Article 164. https://doi.org/10.1007/s44186-026-00567-7

Image Credits: AI Generated

DOI: 10.1007/s44186-026-00567-7

Keywords: laparoscopy, simulation training, 3D visualization, 2D visualization, stereopsis, depth perception, surgical education, box trainer, randomized cross-over trial, low-cost simulator, novice trainees, psychomotor skills

The research, led by Shu Zhang of Huazhong Agricultural University together with Jiapeng Chen, Huan Shen, Quan Liu, and corresponding author Jing Xie, addresses a problem that has long plagued practitioners: labeled stereo data is scarce, expensive, and often unavailable for the specific environments in which a vision system must operate. Stereo matching works by comparing two images captured from slightly offset cameras and computing disparity, the horizontal shift of corresponding pixels, which is inversely proportional to depth. Networks trained on benchmark datasets such as Scene Flow learn this correspondence task superbly, but the statistical properties of those images differ from those of new deployment scenes in lighting, texture, camera geometry, and noise characteristics. This distribution mismatch, known as domain shift, degrades disparity predictions precisely in the regions that matter most for safety, such as thin structures, reflective surfaces, and textureless expanses.

Previous approaches to adaptation have taken several routes. Some methods fine-tune on the target domain using self-supervised photometric losses, exploiting the fact that a correct disparity map should allow one stereo view to be reconstructed from the other. Others rely on domain translation, synthesizing target-like images, or on confidence-based selection of reliable predictions. The new framework departs from these strategies by drawing on an entirely different source of supervision: the recent generation of monocular depth foundation models, such as Depth Anything, Metric3D, and UniDepth, which have been trained on massive unlabeled image collections and generalize impressively across scenes. Rather than ignoring this wealth of relative depth knowledge, the authors harness it as weak ordinal supervision for the stereo network.

The key insight lies in how that supervision is used. Converting monocular depth predictions directly into metric disparity labels is fraught with difficulty. Monocular depth estimates suffer from scale ambiguity, meaning the model may predict depths that are globally stretched or compressed relative to true geometry. Moreover, when depth is expressed in inverse form, as is common in disparity-based formulations, errors are amplified for distant objects, where small absolute depth errors translate into large inverse-depth errors. The team sidesteps both pitfalls by using only relative depth ordering. Their proposed Relative Disparity Loss enforces a simple ordinal constraint: whenever the monocular model judges one pixel to be closer than another, the stereo network is encouraged to predict a larger disparity for that pixel. Because ordinal relationships are invariant to scale and far more robust than absolute values, this constraint transfers reliable geometric knowledge across domains without importing the monocular model’s metric errors.

Ordinal cues alone, however, cannot guarantee accurate and stable adaptation, so the framework adds a consistency-aware learning mechanism with two complementary components. The first is multi-resolution prediction consistency. The stereo network’s predictions are computed at multiple input scales, and the variance across these scales serves as a pixel-wise reliability estimate. Pixels whose disparity predictions fluctuate wildly when the image is resized are deemed unstable and down-weighted in the training loss, while pixels that remain consistent across scales are trusted more heavily. This automatic reliability weighting protects the adaptation process from being corrupted by unreliable regions such as occlusions, sky, and reflective surfaces, which are precisely the areas where naive self-supervision tends to fail.

The second component is feature-level stereo consistency. Instead of constraining only the final disparity maps, the method enforces alignment between left and right image features after warping by the predicted disparity. This feature-space constraint gives the network a richer learning signal than photometric reconstruction alone, because it operates on learned representations that encode semantic and geometric structure rather than raw pixel intensities. As a result, the model becomes markedly more robust under illumination variations, one of the most common causes of performance collapse when synthetic-trained models meet the real world. Importantly, both the monocular depth model and the multi-resolution prediction machinery are used only during training; at inference time, the adapted stereo network runs as a standard, efficient two-view model with no additional computational burden.

The experimental evaluation spans three of the most demanding benchmarks in stereo vision: KITTI, with its autonomous-driving imagery; Middlebury, featuring high-resolution indoor scenes with subpixel-accurate ground truth; and ETH3D, which includes challenging textureless and outdoor environments. Across these targets, the framework achieves competitive or superior domain adaptation performance compared with existing self-adaptive methods, all without ever touching real-world ground-truth disparity during training. Qualitative results are particularly striking in the regions that have historically defeated stereo algorithms. In textureless walls where correspondence matching has no traction, on specular and reflective surfaces where photometric assumptions break down, and around thin structures such as poles and branches where disparity discontinuities are sharp, the adapted network produces visibly cleaner and more coherent depth maps than baselines.

The significance of this work extends beyond a single benchmark leaderboard. It demonstrates a practical recipe for combining the generalization strength of large monocular foundation models with the metric precision of stereo geometry, using the former to supervise the latter in a way that respects the strengths and weaknesses of each. Ordinal supervision sidesteps scale ambiguity, consistency weighting filters out unreliable pixels, and feature alignment hardens the network against lighting shifts. For industries deploying 3D vision, from agricultural robotics to autonomous navigation and industrial inspection, the approach promises models that can be adapted to a new site using only unlabeled stereo pairs captured on location, dramatically reducing the cost and time of deployment.

The study was supported by the National Natural Science Foundation of China under grant 42271357 and the Biological Breeding-National Science and Technology Major Project under grant 2023ZD04029. As monocular depth foundation models continue to improve, the framework’s reliance on relative depth ordering positions it to benefit automatically from future advances, since better ordinal predictions will yield stronger adaptation signals. The work points toward a future in which the divide between benchmark excellence and field reliability finally narrows, allowing stereo vision systems to earn their impressive numbers where it counts: in the unpredictable, uncontrolled, and unlabeled real world.

Subject of Research: Self-supervised domain-adaptive stereo matching using monocular depth cues and consistency-aware learning

Article Title: Domain adaptive stereo matching with consistency-aware learning and monocular depth cues

Article References: Zhang, S., Chen, J., Shen, H., Liu, Q., & Xie, J. (2026). Domain adaptive stereo matching with consistency-aware learning and monocular depth cues. International Journal of Machine Learning and Cybernetics, 17(9), Article 459. https://doi.org/10.1007/s13042-026-03295-y

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03295-y

Keywords: stereo matching, domain adaptation, monocular depth estimation, consistency-aware learning, self-supervised learning, disparity estimation, depth perception, computer vision, KITTI, Middlebury, ETH3D, deep learning

The phenomenon at the heart of the research is known as visual capture, or more specifically in the depth dimension, the proximity image effect. First described in the acoustics literature in the late 1960s, the proximity image effect refers to the tendency of listeners to perceive a sound as originating closer to a plausible visual source than it actually is. When a ventriloquist’s puppet’s mouth moves, the audience hears the voice coming from the puppet, not the performer. In laboratory settings, when a sound is emitted slightly behind or in front of a visible object, listeners systematically misjudge its location, biasing it toward the object. Decades of work, including influential Bayesian models of cue integration, have explained this as the brain weighing visual information more heavily than auditory information for spatial localization, particularly along the depth axis where hearing is comparatively imprecise. Vision dominates because it is typically the more reliable sense for spatial tasks, and the brain behaves like a near-optimal statistical integrator, combining sensory estimates weighted by their respective reliabilities.

What remained unclear, however, was whether these principles, derived mostly from experiments in real rooms with real objects, generalize to virtual environments. Virtual reality is not a neutral replica of the physical world. Decades of research have documented a persistent “distance compression” effect in head-mounted displays: people consistently judge objects in VR as closer, often around 74 to 90 percent of their actual distance, even in modern high-resolution consumer headsets. If the visual representation of space is systematically shrunk, the question becomes how the auditory system, which in these experiments received real, physically accurate sound, interacts with a visual system operating on distorted coordinates. The Utah team, led by Maggie K. McCracken together with Macy E. Rowland, Hunter C. Finney, Jeanine K. Stefanucci, and Sarah H. Creem-Regehr, designed two experiments to answer exactly this question.

The experimental logic was elegant. Participants wore an immersive head-mounted display in a real laboratory room while hearing sounds from an actual, physical sound source positioned at various depths. At the same time, they saw a virtual object rendered at a specified depth in the virtual scene. Their task was a perceptual matching judgment: they indicated when they perceived the real-world sound source to be aligned in depth with the virtual object. If no visual capture occurred, their judgments should track the true physical co-location of sound and virtual object. If visual capture occurred, their points of subjective alignment should be biased toward the visible object, and the magnitude and direction of that bias would reveal how the brain was combining the two spatial estimates. Crucially, this setup created a genuine cross-reality conflict: the ears received physically veridical information about the real world, while the eyes received rendered information about a virtual object occupying a different depth position.

The results of the first experiment delivered a double surprise. Visual capture did reliably occur, confirming that the multisensory machinery at work in physical rooms also operates when one of the signals belongs to a synthetic environment. But the asymmetry of the effect ran opposite to what real-world studies had established. In classic experiments, capture was stronger when the sound originated behind the visual object, presumably because listeners are more uncertain about sounds located behind opaque surfaces and therefore lean more heavily on vision. In the Utah study, capture was stronger when the sound originated in front of the virtual object rather than behind it. At first glance, this reversal looked like a failure of the classical model, a suggestion that the ventriloquist effect might not behave the same way in VR at all.

The second experiment resolved the puzzle. The team replicated the reversed asymmetry pattern, but they also added an independent measurement of the participants’ visual distance perception using a blind-walking task, a gold-standard procedure in which participants view a target and then walk to it without vision, allowing researchers to infer perceived distance from walked distance. The measurements confirmed what the VR literature had long predicted: participants significantly underestimated the visual distances of virtual objects. When the researchers reanalyzed the audiovisual alignment judgments not relative to the physical distance of the virtual objects, but relative to each participant’s perceived, compressed visual distance, the expected real-world symmetry reemerged. In other words, the brain’s integration rule had not changed; the coordinate system had. Once the analysis was anchored to what participants actually perceived visually rather than where the virtual object “really” was in the rendered scene, the classic pattern of capture fell back into place.

This reframing carries significant theoretical weight. It implies that multisensory integration operates on perceived space, not physical space. The brain does not have access to the ground-truth coordinates programmed into the virtual environment; it only has access to the noisy perceptual estimates generated by the visual and auditory systems. If the visual estimate is compressed, then the point at which sound and vision appear co-located shifts accordingly, and the pattern of capture shifts with it. The findings thus demonstrate that core principles of multisensory integration, including visual dominance in depth and the statistical weighting of cues by reliability, are preserved in virtual reality, but that VR-specific biases can alter, and even reverse, the direction of well-established effects when measured against physical reality. This is a powerful demonstration that the same computational rules apply across real and virtual worlds, provided researchers account for how those worlds are actually perceived.

The methodological implications are substantial for anyone building or studying virtual environments. Developers of spatial audio for VR often assume that rendering sound at a physically correct position relative to a visually rendered object will produce the intended percept. This study shows that the perceived relationship between a real sound and a virtual object can deviate in systematic, predictable ways from the physical arrangement, and that these deviations are driven by the visual system’s compressed representation of depth. Applications that depend on precise audiovisual alignment, such as virtual training simulations, spatial audio design for games and cinematic VR, accessibility tools that use sound to guide attention to visual targets, and clinical or research paradigms involving multisensory integration, may all need to compensate for distance compression when placing sounds relative to visual content. Conversely, the effect could be exploited deliberately: if visual capture reliably pulls perceived sound location toward virtual objects, designers could use visual anchors to sharpen the perceived localization of audio without changing the audio itself.

The work also contributes to a deeper scientific conversation about why distance compression occurs in head-mounted displays and how it interacts with other perceptual processes. Hypotheses have ranged from limited field of view and display resolution to the absence of accurate accommodative and vergence cues, the influence of the virtual environment’s layout, and the lack of natural body-based scaling information. By showing that distance compression propagates into the domain of multisensory perception, the study adds a new consequence to the list: the distortion does not merely make virtual rooms feel smaller, it reshapes how the brain binds together information from different senses within those rooms. That makes distance perception in VR not just a curiosity of display technology but a foundational parameter that conditions virtually every spatial judgment a user makes.

The research, conducted at the University of Utah’s Department of Psychology and approved by the university’s Institutional Review Board, was supported in part by a National Science Foundation Graduate Research Fellowship. Trial-level data and analysis scripts for both experiments have been made publicly available on the Open Science Framework, reflecting a growing commitment to transparency in perception science. As head-mounted displays become everyday consumer technology and spatial audio becomes a standard feature of digital experiences, understanding how the brain reconciles real sound with virtual sight will matter far beyond the laboratory. This study offers both a caution, that virtual worlds distort space in ways that silently reshape multisensory perception, and a reassurance, that the brain’s integration rules remain coherent, lawful, and ultimately predictable even in worlds that do not physically exist.