Answers are grounded in this paper and cite their sources.
Transect surveys are a standard sampling method in diver-led underwater environmental assessment, but, to our knowledge, no prior public dataset treats the transect tape as the primary target for autonomous robotic following. We present a 12,742-image multi-source underwater dataset with polygon annotations, spanning controlled pool recordings and multiple outdoor sources. A trained segmentation model is evaluated on unseen acquisition sources, alongside a separate benchmark of zero-shot methods. We also introduce a model-agnostic dynamic region-of-interest (ROI) method that uses previous detections to focus segmentation and expands the search after detection loss. The resulting tape position and orientation estimates support robotic following. Validation includes video replays, pool trials with motion capture, and open-water deployments. YOLOv8n-seg achieves mask mAP@50 of 0.8135 on a held-out video source, although performance varies across acquisition sources. In three open-water runs at a previously unseen site, the system provided tape position and orientation estimates for 87.8% of the combined run time. The dataset supports reproducible evaluation of transect perception, and the field trials provide evidence of generalisation to a new underwater environment.
No takes yet. Share an insight, caveat, or question.
Sultan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: