Benchmarking Real-Time DETR Detectors for Multibeam Forward-Looking Sonar: DEIMv2, Detector-Specific Test-Time Fusion, and Domain Shift Analysis on UATD
Year: 2026
Type: preprint
Abstract: Multibeam forward-looking sonar (MFLS) target detection remains challenging because of low spatial resolution, acoustic speckle, and distributional differences across survey conditions. This study benchmarks DEIMv2, a DINOv3-based real-time DETR-family detector, on the Underwater Acoustic Target Detection (UATD) dataset under a specified, reproducible dual-test-split protocol. We compare DEIMv2 with RT-DETR-R18, D-FINE-S, DEIM v1-S, and YOLOv8-S under a common dataset split, input resolution, training budget, augmentation schedule, and evaluation pipeline, while retaining model-specific initialization and official optimization details (Section 4.2), and evaluate weighted boxes fusion (WBF) test-time augmentation (TTA). Without TTA, DEIMv2-S achieved 79.9% and 82.3% [email protected] on Test_1 and Test_2, respectively, with mAP@[.5:.95] of 37.7% and 38.2%. Its mAP@[.5:.95] was comparable to the other query-based detectors, whereas its [email protected] was lower than that of the earlier DETR baselines. For DEIMv2-S, six-view WBF TTA improved [email protected] by 1.4 and 2.1 percentage points on Test_1 and Test_2, respectively; however, the same fusion recipe was neutral or detrimental for RT-DETR-R18 and D-FINE-S and reduced AR@100. Thus, the benefit is detector-specific and should not be assumed for recall-critical applications. We further find a 12–16-point gap between validation and test performance, accompanied by shifts in class and object-size distributions; histogram matching did not remove this gap; the evidence is consistent with substantial distributional shift between the training/validation and test partitions, whose specific causes cannot be identified from the available metadata. The study provides a reproducible reference for current DETR-family detectors on MFLS imagery and identifies domain adaptation and detector-aware test-time fusion as priorities for future work.
Source: Preprints.org
Language: English
Subfield: Oceanography
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213754151