A Scientific Reports paper published October 18 describes GDH-Det, which uses conditional image-degradation simulation and hard-sample selection to train an underwater object detector. Across four public datasets, the authors report mAP@0.5 of 85.3%, 79.6%, 66.2% and 84.7%. These are benchmark results, not evidence of performance in deployed underwater vehicles.
GLOBAL IMAGING BRIEF
Analysis: The work gives marine-vision developers a reproducible way to test robustness to turbidity, blur and uneven illumination; cross-site validation and hardware-constrained inference remain separate engineering questions.
Original source
Original peer-reviewed paper published October 18, 2026. Benchmark figures are the authors' reported results.