cs.CVApr 14, 2026
Save4th Workshop on Maritime Computer Vision (MaCVi): Challenge Overview
Abstract
The 4th Workshop on Maritime Computer Vision (MaCVi) is organized as part of CVPR 2026. This edition features five benchmark challenges with emphasis on both predictive accuracy and embedded real-time feasibility. This report summarizes the MaCVi 2026 challenge setup, evaluation protocols, datasets, and benchmark tracks, and presents quantitative results, qualitative comparisons, and cross-challenge analyses of emerging method trends. We also include technical reports from top-performing teams to highlight practical design choices and lessons learned across the benchmark suite. Datasets, leaderboards, and challenge resources are available at https://macvi.org/workshop/cvpr26.
Figures & tables
Figure 1: Qualitative detection examples from the Vision-to-Chart Data Association challenge. Each panel shows the overlay of predicted and ground-truth bounding boxes and predicted index of associated chart marker.
Table 1: Results for the Vision-to-Chart Data Association challenge on the held out test split. Ranking medals indicate final placement. The baseline (gray) is the organizer-provided DETR-style fusion transformer. Method names are derived from the submitted technical reports.
Table 2: Results for the Thermal Object Detection challenge. Ranking medals indicate final placement. The baseline (gray) is the organizer-provided Faster R-CNN with ResNet-50. Self-reported inference speed and hardware are included. Backbone and detection-level pretraining datasets are reported separately (IN = ImageNet-1K, IN-22K = ImageNet-22K, LVD = Large Vision Dataset).
Figure 2: Per-class and per-size AP breakdown for the Thermal Object Detection challenge. Left: AP for vessel vs. navigational object classes. Right: AP at COCO size thresholds (small / medium / large). Vessel detection is consistently easier than navigational object detection across all methods. Small object AP remains the main bottleneck.
Figure 3: Qualitative detection examples from the Thermal Object Detection challenge. Each row shows one test image with ground truth (green) and predictions from the top three methods and baseline. Columns left to right: ground truth, 1st place, 2nd place, 3rd place, baseline. Predictions are filtered at confidence . V = vessel, N = navigational object.
Figure 4: Precision–Recall curves for the Thermal Object Detection challenge. Curves are shown for each method at IoU=0.50.
Table 3: Results for the USV-based Panoptic Segmentation Challenge measured in overall panoptic quallity (PQ) and separate for thing and stuff classes. Competing methods are compared to a baseline Mask2Former model [ 5 ] and the top-performing PanSR method [ 59 ] .
Figure 5: Panoptic quality stratified according to scene attributes.
Figure 6: Dynamic obstacle detection rate stratified according to object size.
Figure 7: Qualitative results of the top-performing methods in the LaRS panoptic segmentation challenge.
Table 4: Overview of the submissions for the USV-based Embedded Obstacle Segmentation challenge. For comparison, last year’s winning method is included inline in gray and is not assigned placements. The best results among current non-gray entries are denoted in bold.
Figure 8: Detection rate and FP as a function of obstacle size for embedded segmentation methods. Only the three winning and the best-performing method from last year’s challenge are shown.
Figure 9: Detection F1 under challenging visual conditions for embedded segmentation methods. Only the three winning and the best-performing method from last year’s challenge are shown.
Figure 10: Detection F1 across scene environments and reflection levels for embedded segmentation methods showing the three winning and the best-performing method from last year’s challenge.
Table 5: Results for the USV-based Multimodal Segmentation Challenge. We report , validation/test mIoU, and dynamic-obstacle IoU.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | PQ | SQ | RQ |
| Mask2Former + DINOv3-L | 54.8 | 75.8 | 63.5 |
| Mask2Former + DINOv3-H+ | 55.8 | 75.5 | 65.4 |
Table 6: Panoptic segmentation results on the LARS validation set.
Figure 11: Training pipeline. WS = WaterScenes, RB = ROSEBUD, MA = MULTIAQUA_night.
Figure 12: Overview of GatedMemorySAM. Each modality is encoded by SAM2 with Soft-MoE LoRA, scored by the CrossModalFusionHead, and fused via memory modulation and weighted mask fusion.