Cross-Domain Robustness of CNN vs Transformer Detectors for Underwater Object Detection

Abstract

Underwater object detection is limited by two compounding problems: strong domain shift across datasets caused by variable water conditions (attenuation, haze, turbidity, and blur), and label noise arising from missed annotations in low-visibility images. We systematically compare a CNN detector (YOLOv8) and a Transformer detector (DETR) under cross-dataset transfer, physically motivated underwater corruptions, and noisy supervision, and propose a training strategy that improves cross-domain robustness without re-annotation. Our method combines (i) a raw↔enhanced feature-consistency loss using a cheap CLAHE-based underwater enhancer as an auxiliary view, and (ii) a noise-aware supervision reweighting that down-weights samples whose raw and enhanced features strongly disagree. Experiments on DUO and RUOD show three main findings. First, YOLO achieves higher absolute performance than DETR in-domain and cross-domain, but relative robustness is direction-dependent: DETR retains more in-domain mAP under DUO→RUOD shift (66.2% vs. 56.4%), whereas YOLO retains more under RUOD→DUO (40.3% vs. 29.9%). Second, haze and turbidity are the most destructive corruptions, while blur is the least harmful. Third, the proposed method is beneficial but detector- and direction-dependent: for DETR, the full method improves DUO→RUOD cross-domain mAP from 0.285 to 0.294 and AP50 from 0.608 to 0.614, whereas for YOLO the best variant is weight only on DUO→RUOD and full on RUOD→DUO.

Limitations & future work

• DETR was trained for only 30 epochs at batch size 2; longer training and batch size ≥ 8 should make the noise-aware weighting meaningful for DETR.
• Next steps: stronger transformers such as RT-DETR, physics-informed enhancement models, and domain-specific augmentation to reduce CNN texture bias.
• Evaluation covers 4 shared classes. Re-annotating the RUOD test set with enhanced images as reference would give a cleaner upper bound on true cross-domain performance.

Code

Citation

Copy@techreport{almulla2026underwater,
  title       = {Cross-Domain Robustness of CNN vs Transformer Detectors
                 for Underwater Object Detection},
  author      = {Almulla, Buthaina and Alshehhi, Abdulla and Yaqoob, Shamma},
  institution = {Mohamed bin Zayed University of Artificial Intelligence},
  address     = {Abu Dhabi, United Arab Emirates},
  year        = {2026}
}