Underwater archaeology stops relying on color alone: adding a depth stream lifts WRF mAP75 to 88.0%

Automated detection for underwater cultural heritage now has a dedicated multimodal benchmark whose targets are shipwreck structures, pottery jars, porcelain, anchors and rudders rather than ecological objects such as sea cucumbers and fish. Earlier underwater benchmarks such as URPC2020, DUO and RUOD mainly contain benthic organisms, UIIS covers fish and coral reefs, and TrashCAN targets marine debris, so their category definitions and scene composition do not match archaeological objects. WRF collects images from multiple deep-sea archaeological sites in the Gulf of Mexico, the Black Sea and the Mediterranean; after manual deduplication and quality screening it retains 4,732 RGB images with 17,059 annotated instances across 21 artifact and ship-component categories, with a long-tailed class distribution.

The detector reads color appearance and depth geometry together and adjusts how much it trusts each according to the degree of imaging degradation. RGB-only detectors lose color, texture and boundary cues under weak illumination, scattering and substantial optical degradation, while earlier color-correction pipelines generally overlook how color shift varies non-uniformly with propagation distance. On WRF, UCH-Net reaches 94.9% precision, 89.9% recall, 92.1% F1, 94.4% mAP50, 88.0% mAP75 and 77.5% mAP; LFN-YOLO has the highest precision at 95.2%, 0.3 percentage points above UCH-Net.

Under a stricter localization threshold the advantage is larger than under the looser one, which matters for archaeological workflows that need to outline an object's extent rather than only decide whether it is present. Against BSR5-YOLO, the closest competitor in overall mAP, UCH-Net is 0.5 percentage points higher in mAP but 1.2 percentage points higher in mAP75. On the public RUOD dataset, UCH-Net reaches 87.3% precision, 80.6% recall, 83.7% F1, 87.3% mAP50, 72.7% mAP75 and 66.0% mAP, with LFN-YOLO holding the highest recall at 80.9%.

The demonstrated scope is detection evaluation on the WRF and RUOD datasets, with WRF split 8:2 into training and test in temporal order and RUOD following its official split; the authors propose future work on measured depth, confidence-aware fusion and site-independent splits to test cross-site generalization.

Depth maps are estimated from RGB images by Depth Anything V2 and may be affected by underwater color distortion, scattering, weak textures and domain differences between underwater scenes and that model's training data, so they are better understood as auxiliary geometric priors than as physical distance measurements; WRF currently uses a temporally ordered image-level 8:2 split, its class distribution is long-tailed, and experiments rely mainly on single runs; UCH-Net has 48.8 M parameters but needs 143.8 GFLOPs per forward pass, higher than most lightweight YOLO-based detectors.