Marine roboticists have developed an adaptive object detection framework named Hy-YOLO that significantly enhances visual perception for unmanned surface vehicles operating across diverse and unpredictable ocean conditions. Published in Frontiers in Robotics and AI by researchers from China University of Geosciences, Zhejiang University, and Shanghai Maritime University, the experimental study demonstrates that decoupling structural deformation modeling from numerical bounding box regression resolves the performance compromises that traditionally hamper general-purpose vision algorithms at sea. Deploying autonomous surface craft requires visual detectors to identify ships oscillating in turbulent waves while simultaneously pinpointing tiny floating hazards across open water expanses. Rather than relying on a rigid, one-size-fits-all deep learning model, the authors established an empirical selection rule that routes visual data through specialized algorithmic paths tailored to immediate environmental characteristics, providing an adaptable architecture capable of sustaining real-time processing speeds.
The fundamental dilemma confronting maritime perception stems from two conflicting imaging realities that standard convolutional neural networks struggle to reconcile. In turbulent nearshore waters, surge waves induce severe six-degree-of-freedom motion in surface craft, creating significant non-rigid geometric distortions, wake turbulence, and intense surface glint that deceive fixed convolutional grids. Conversely, open-sea navigation presents expansive vistas where objects such as navigation buoys or marine debris occupy fewer than one percent of an image’s total pixels, rendering localization exceptionally vulnerable to minor coordinate regression errors. The research team identified what they term a stacking paradox in contemporary maritime detection research: indiscriminately layering multiple complex modules, such as spatial attention, deformable kernels, and heavy loss functions, frequently yields diminishing returns or degrades detection accuracy in straightforward scenarios. The researchers hypothesized that disparate maritime environments require distinct inductive biases, specifically demanding dynamic structural alignment for turbulent wave states and strict geometric boundary constraints for sparse small objects.
To resolve these conflicting challenges, the investigators formulated the Hy-YOLO framework around a dual-component architecture consisting of a universal attention baseline coupled with scenario-specialized branches. The foundation of this system is a serial composite attention mechanism placed at the termination of the feature extraction backbone network. The authors arranged a Convolutional Block Attention Module, known as CBAM, in direct sequence with an Efficient Channel Attention module, abbreviated as ECA. The preliminary CBAM component sequentially infers spatial and channel attention maps using combined average-pooling and max-pooling operations, suppressing broad swaths of background sea clutter. The subsequent ECA component processes the filtered representations using a one-dimensional convolution with an adaptive kernel size to capture localized inter-channel dependencies. By operating as a learnable spectral filter, this composite attention mechanism cleans raw visual features prior to scenario routing, reducing the false positive rate of background water being misclassified as vessels by approximately fifty-four percent.
Building upon this universal attention base, the investigators designed the Hy-YOLO-G variant specifically to overcome structural deformation in turbulent maritime environments. Standard convolution operations rely on rigid rectangular sampling grids that fail to capture skewed hulls or occluded superstructures during heavy wave pitching. To solve this limitation, Hy-YOLO-G integrates third-generation Deformable Convolution, designated as DCNv3, into the neck of the detection network. The DCNv3 operator utilizes shared projection weights alongside a multi-group modulation mechanism, dividing convolutional channels into distinct aggregation groups to expand feature diversity. For every spatial sampling coordinate, the module dynamically calculates both continuous geometric offsets and scalar modulation values. This architecture permits the receptive field to reshape itself organically across distorted target contours, functioning akin to an amoeba that molds its sampling points directly over vessel boundaries rather than absorbing surrounding water surface noise. Consequently, target features undergo precise geometric alignment prior to subsequent bounding box prediction layers.
For tranquil open-sea regimes containing tiny, rigid targets, the authors introduced the Hy-YOLO-L variant, which substitutes complex structural modification with advanced loss function engineering. Because minute targets possess negligible pixel footprints, minor coordinate discrepancies precipitate steep drops in intersection over union metrics, while gradients remain chronically vulnerable to swamping by vast, uninformative marine backgrounds. Hy-YOLO-L resolves this drift by superimposing a hybrid loss formulation onto the baseline Complete Intersection over Union objective throughout training. This formulation merges SCYLLA Intersection over Union, termed SIoU, with Focal Efficient Intersection over Union, termed Focal-EIoU. The SIoU component introduces an explicit angle cost that penalizes angular deviations between predicted coordinates and ground-truth boxes, compelling predicted boundaries to align directly along Cartesian axes. Simultaneously, Focal-EIoU weights bounding box errors according to sample difficulty, actively penalizing easily classified water background regions while amplifying gradients from challenging hard positive targets. Hyperparameter testing established optimal balancing coefficients of one-half for SIoU and one for Focal-EIoU, stabilizing fine-grained regression without altering backbone parameters.
Empirical validation of the framework utilized two contrasting benchmark datasets evaluated on an identical computational platform featuring an Intel Core i7-13700H processor and an NVIDIA GeForce RTX 4060 Laptop GPU with eight gigabytes of memory. To evaluate turbulent dynamic scenes, the authors employed the Water Surface Object Detection Dataset, known as WSODD, comprising 21,911 annotated instances across fourteen categories, including boats, buoys, platforms, and personnel amid severe wave action and white water foam. For open-water fine localization, the team tested the Flow dataset, containing 3,248 labeled floating bottles exhibiting an average pixel coverage under one percent in calm water. Models underwent eighty training epochs using stochastic gradient descent with an initial learning rate of 0.01, a batch size of four, momentum of 0.937, and weight decay of 0.0005. Data augmentation incorporated mosaic construction, random scaling, translation, horizontal flipping, and color jittering, with mosaic augmentation disabled across the terminal ten epochs to secure stable numerical convergence across all baseline and modified detectors.
Experimental assessments on the turbulent WSODD benchmark verified the distinct advantage of the structural adaptation branch. The standard YOLOv8s baseline posted a mean average precision at fifty percent intersection over union of 0.6221. Embedding the universal CBAM and ECA attention frontend lifted precision to 0.7271 while sustaining mean average precision at 0.6222. Incorporating the DCNv3 deformable convolution within Hy-YOLO-G boosted mean average precision to 0.6371, reflecting a 1.5 percentage point enhancement, while driving strict multi-threshold precision across IoU thresholds from 0.5 to 0.95 to 0.3292. Crucially, the researchers conducted a deliberate control trial by forcing the localization-focused Hy-YOLO-L configuration onto the WSODD dataset. Rather than achieving improvements, Hy-YOLO-L registered a reduced mean average precision of 0.6113, underperforming the initial baseline by 1.1 percentage points. This failure proved that mathematical loss constraints applied at the bounding box regression head cannot rectify spatial error when convolutional features suffer from uncorrected geometric misalignment caused by dynamic wave perturbations.
Evaluations conducted on the Flow dataset demonstrated an inverted performance trajectory, confirming the scenario-dependent nature of vision optimization. On this benchmark, the standard YOLOv8s baseline recorded a mean average precision at fifty percent intersection over union of 0.8754. Simply introducing the universal attention baseline yielded an increase to 0.9266, underscoring the efficacy of attention filters in eliminating sporadic water reflections. Expanding upon this foundation with the hybrid SIoU and Focal-EIoU loss formulation of Hy-YOLO-L delivered an overall mean average precision of 0.9173, reflecting a 4.2 percentage point gain over the unassisted baseline. Conversely, inserting the DCNv3 deformable convolution into this small-target context precipitated a notable performance decline, dropping precision to 0.8993. The authors observed that because small floating targets remain structurally rigid, the supplementary parameter overhead and offset prediction mechanisms of deformable convolutions induce overfitting and spatial fragmentation. This sharp divergence directly validated the stacking paradox, proving that structural deformation modeling is counterproductive when applied to rigid, minute maritime objects.
To demonstrate that their empirical findings were not artifacts of the YOLOv8 architecture, the researchers evaluated the framework on an open-source YOLOv13 baseline possessing an anchor-free design and an A2C2f backbone. On WSODD, the YOLOv13 baseline began at a lower precision of 0.5732, but adding the Hy-YOLO-G deformable pathway elevated accuracy to 0.6012. On Flow, the YOLOv13 baseline reached 0.8484, which advanced to 0.8913 following Hy-YOLO-L integration. Concurrently, the team examined two multi-scale inference strategies designed to capture minute targets: high-resolution input at 1,536 by 1,536 pixels versus sliding-window image slicing. While image slicing elevated precision on the clean Flow dataset from 0.875 to 0.941, it triggered severe accuracy collapse on WSODD, sinking precision to 0.619 below the initial baseline. The authors determined that spatial slicing shears large hulls and severs global semantic cues, such as horizon continuity and wake lines, generating numerous false alarms. Conversely, high-resolution processing maintained global contextual integrity, lifting accuracy across both benchmarks to 0.6491 on WSODD and 0.9455 on Flow.
Robustness testing under simulated marine optical degradation reinforced these scenario-driven findings. Under synthetic atmospheric scattering fog on Flow, Hy-YOLO-L maintained an mAP of 0.7808, while Hy-YOLO-G resisted intense wave-crest Gaussian bloom glare on WSODD with an mAP of 0.5505. In zero-shot transfer trials on the 7,000-image SeaShips benchmark, Hy-YOLO-G yielded a 7.7 percent relative accuracy improvement over baseline on large ships, whereas Hy-YOLO-L achieved a 257 percent relative boost in localizing novel pixel clusters. Computational profiling on the RTX 4060 GPU confirmed real-time capability, with Hy-YOLO-G and Hy-YOLO-L operating at 67.6 and 66.6 frames per second at standard resolution. The authors proposed deploying the efficient Hy-YOLO-L branch during open-sea cruising and engaging Hy-YOLO-G during nearshore maneuvering. Remaining limitations include missed detections of personnel aboard large vessels and targets occupying under five pixels, motivating planned research into multi-spectral infrared fusion and automated branch switching.
Subject of Research: Scenario-guided dual-path object detection framework for maritime unmanned surface vehicles
Article Title: A dual-path scenario-guided framework for maritime object detection under heterogeneous sea conditions
Article References: Li, Y., Zhao, H., Kong, Y., Wei, L., Wu, D., & Zhao, J. (2026). A dual-path scenario-guided framework for maritime object detection under heterogeneous sea conditions. Frontiers in Robotics and AI, 13, 1865853. https://doi.org/10.3389/frobt.2026.1865853
Image Credits: AI Generated
DOI: 10.3389/frobt.2026.1865853
Keywords: maritime object detection, unmanned surface vehicles, Hy-YOLO, YOLOv8, YOLOv13, deformable convolutions, DCNv3, composite attention mechanism, CBAM, ECA module, SIoU loss, Focal-EIoU loss, WSODD dataset, Flow dataset, bounding box regression, sea state dynamics, specular glare robustness, synthetic fog testing
Cite Scienmag News
Denise Maddox. (November 17, 2026). Scenario-Guided Vision Framework Improves Object Detection Accuracy for Unmanned Surface Vessels. Scienmag. https://scienmag.com/scenario-guided-vision-framework-improves-object-detection-accuracy-for-unmanned-surface-vessels/
Denise Maddox. "Scenario-Guided Vision Framework Improves Object Detection Accuracy for Unmanned Surface Vessels." Scienmag, 17 November 2026, https://scienmag.com/scenario-guided-vision-framework-improves-object-detection-accuracy-for-unmanned-surface-vessels/. Accessed 17 November 2026.
Denise Maddox. "Scenario-Guided Vision Framework Improves Object Detection Accuracy for Unmanned Surface Vessels." Scienmag. November 17, 2026. https://scienmag.com/scenario-guided-vision-framework-improves-object-detection-accuracy-for-unmanned-surface-vessels/








