AquaMambaTrack: Reliability-First Mamba ReID for Underwater Fish Multi-Object Tracking

Abstract

High-density underwater fish multi-object tracking remains difficult because occlusion, small targets, background similarity, and noisy observations destabilize identity evidence. We study a reliability-first temporal identity memory in which a successful association does not automatically authorize an identity-memory update. AquaMambaTrack (AMT) decouples geometric/motion-state updates from admission to an online per-track FIFO: only observations that satisfy stage, geometry, appearance, and crowd-reliability checks enter the history aggregated by Mamba. With fixed official MFT25 detections and official TrackEval, AMT-L48 obtains 53.473 HOTA, 41.112 AssA, 59.543 IDF1, 85.265 MOTA, and 507 IDSW. In the write-policy ladder, unfiltered FIFO writing gives 47.973 HOTA and 33.189 AssA, whereas the complete policy gives 53.473 and 41.112; the GT-evaluable history-contamination rate decreases from 27.89% to 14.35%. Online track-conditioned history exceeds detection-linked history without fallback by 5.933 HOTA and the documented T = 1 fallback path by 4.210 HOTA. Across seeds 0, 1, and 2, L = 48 gives 52.662 ± 0.769 HOTA, and Mamba has the highest mean association metrics among the evaluated encoders in the same pipeline. The gain is not universal: at the common detection-score threshold of 0.15, SU-T has fewer raw ID switches (381) on MFT25; in the internal cross-scene case study, AMT leads HOTA, DetA, and MOTA but not AssA or IDF1, and has the most raw ID switches (82). Under precomputed detections, the cold-cache tracking pipeline runs at 30.33 ± 2.72 FPS.