Vizualno sledenje objektom na podlagi pomnilnika z upoštevanjem distraktorjev

The Segment Anything Model 2 (SAM2) has recently gained significant attention for its strong performance in segmentation tasks, achieving leading results on numerous benchmarks. However, while SAM2 serves as a powerful foundation for video segmentation, its architecture is not fully optimized for visual object tracking. Specifically, we identify distractors as a key limitation that decreases tracking robustness over time. In this thesis, we refine SAM2’s memory mechanism and propose DAM4SAM: a distractor-aware drop-in memory model for SAM2 paired with an introspection-based management. The memory design successfully reduces tracking drifts to the distractors and improves redetection capability after object occlusion. To fascilate deeper analysis of tracking in the presence of distractors, we construct DiDi, a Distractor-Distilled dataset. DAM4SAM outperforms SAM2.1 on thirteen benchmarks and sets new state-of-the-art results on ten. Moreover, the proposed distractor-aware memory improves the recent SAM2-based realtime tracker EfficientTAM by 11% on DiDi, and also matches SAM2.1-L performance on multiple tracking and segmentation benchmarks, demonstrating strong generalization capabilities.