Abstract
Multibeam forward-looking sonar (MFLS) is the primary sensing modality for underwater vehicles operating in turbid water, yet low resolution, acoustic speckle, and domain shift keep detection challenging. On the public Underwater Acoustic Target Detection (UATD) benchmark, validation performance does not reliably predict test performance: a persistent gap of 13.3–17.8 points in mAP@0.5 separates the two. We benchmark DEIMv2, a DETR-family detector on DINOv3 foundation features, against four baselines under one harmonized protocol. Without test-time augmentation (TTA), DEIMv2-S reaches 79.9%/82.3% mAP@0.5 on Test_1/Test_2, trailing earlier DETR generations by 3.3–4.1 and 1.3–2.0 points; the four query-based detectors cluster at 37.3–38.2% strict mAP@[.5:.95], while DEIMv2 converges with one third of the standard RT-DETR training budget and runs near real time per view. Weighted boxes fusion (WBF) TTA lifts Test_2 mAP@0.5 to 84.4% but costs AR@100; a view-deduplicated rule improves strict mAP and AR@100 in all eight model–split cells and is our recommended variant. These gains are detector-dependent—the same query-level recipe is neutral to harmful for the baselines—and the validation–test gap is consistent with an acquisition-condition shift rather than model capacity.