An underwater detector that combines camera images with sonar reported modestly better results than DINO on the RSFusion benchmark. RSFusionDet recorded 76.4 Average Precision, or AP, for RGB detection and 48.6 AP for sonar. Those figures were reported as gains of 0.7 and 1.4 AP points, respectively, over the DINO comparison model.
That gives the paper two related tests. It measures how well the system finds objects in each sensor's data, then separately checks whether detections in the RGB and Sonar views refer to the same object. The evaluation uses standard COCO Average Precision for the two detection tasks. For correspondence, the authors use an object-matching head and loss designed to link identical objects across the modalities.
The benchmark behind the comparison
The RSFusion benchmark is an underwater RGB-Sonar object-detection benchmark with 7,073 paired images covering seven object classes. Its split used 5,658 pairs for training and 1,415 for validation.
All compared models were trained and tested on the RSFusion training and validation sets. The result is a benchmark comparison within the same data split, with RSFusionDet's scores measured against DINO under the study's stated evaluation arrangement.
How the two sensor views are combined
The model's central fusion component is CAFusion. The paper describes it as a cross-attention module based on multi-scale deformable attention for RGB-Sonar feature fusion. In plain language, the module is intended to let the two sensor streams exchange selected feature information at several scales, rather than relying on only one view.
The architecture adds a second mechanism for correspondence. Its object-matching head and loss are designed to establish links between identical RGB and Sonar objects. That matters because the detector is not only reporting whether an object is present; it is also assigning detections across the two modalities.
Where the reported gains appeared
On that matching task, the selected OMHead configuration reported a precision score of 80.9, a recall score of 86.0 and an F1 score of 83.4. The scores were obtained at a filtering threshold of 0.3 and a matching threshold of 0.8. In the paper's terminology, these are the reported precision, recall and combined F1 measures for cross-modal matching.
Component ablations compared the two named parts. The CAFusion ablation gave RGB and sonar AP differences of 0.5 and 1.1 points, while the OMHead configuration reported an F1-Scorematch of 83.4. These comparisons describe score changes among the configurations tested in the benchmark; they do not by themselves show that the same differences would appear under another evaluation setup.
Performance was also separated by scene brightness. Compared with DINO, the reported RGB and sonar AP differences were 0.8 and 1.1 points in light scenes, and 1.0 and 1.2 points in dark scenes. The gap was therefore slightly larger in the dark-scene subgroup for both modalities, although the results remain benchmark-specific.
The range analysis showed a different split. As maximum sonar range increased, the authors reported that RGB improvement relative to DINO increased, while sonar improvement decreased. The supplied result gives the direction of the pattern across range conditions, without a separate numeric value for each condition.
What the benchmark does not answer
The main unanswered issue is scope. Because all compared models were trained and tested on RSFusion's training and validation sets, the reported AP and matching figures describe that benchmark comparison. They do not by themselves establish how the system would perform outside the tested data.
The authors state that the code and datasets are publicly available. The document is an arXiv version-1 preprint dated 26 Aug 2026. For now, the clearest reading is a reported benchmark gain paired with a reported cross-modal matching result, with broader validation left open.
Paper data and sources
Original title: RSFusionDet: Underwater RGB-Sonar Multimodal Object Detection
Authors: Zhuoyan Liu, Yihan Wang, Bo Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text