Mot tillförlitlig objektspårning och segmentering i video med hjälp av SAM2

This thesis enhances the SAM2 framework for video object tracking, focusing on occlusion, background clutter, and long-term target reappearance— scenarios where existing methods often fail. These challenges are both practically relevant and technically complex, making them appropriate for in-depth academic investigation. Two complementary solutions are proposed. First, a trainless method improves tracking without additional training. It combines hierarchical motion estimation—linear prediction with selective non-linear refinement—and an optimized memory mechanism. This approach achieves 75.1% AUC on LaSOT and 62.8% on LaSOText with the large SAM2 model, and shows significant relative gains on smaller variants, indicating strong scalability. Second, a trainable strategy integrates motion prediction with Mamba-based long-term memory modeling. A custom dataset targeting reappearance scenarios is developed to enable supervised training. Although constrained by hardware, initial experiments show modest improvements over the baseline, suggesting potential for further performance with more extensive training. Together, these methods significantly improve SAM2’s robustness in complex video tracking tasks. The project contributes practical solutions that enhance current capabilities while laying groundwork for future advances in scalable, long-term tracking.