OAMVOS: Occlusion-Aware SAM Tracker for MOSE Challenge

A new report presents OAMVOS, an occlusion- and reappearance-aware extension of DAM4SAM, submitted to the 5th PVUW MOSE Track. The method addresses fragility in SAM-based dense trackers under long occlusion, fast motion, viewpoint changes, and distractors, particularly for small objects. It improves memory control via a reliability-aware state machine, branch-based recovery, delayed DRM promotion, and selective memory selection. During stable tracking, the model uses single-path propagation; when confidence drops, it enters ambiguous or recovery mode, maintains candidate branches, and commits memory only after reconfirmation. The approach targets small-object disappearance and reappearance without changing the backbone.