SAM 3 is Meta’s next-generation segmentation model that shifts from geometric, prompt-based segmentation to concept-level understanding through Promptable Concept Segmentation, enabling open-vocabulary instance detection via text or visual exemplars.

The Segment Anything Model 2 (SAM-2) transformed video object segmentation with its memory-based architecture for sequential frames. However, it struggles with occlusions and error propagation. Derivative models like SAMURAI and SAM2Long address these issues by integrating advanced memory and motion-aware mechanisms, improving segmentation accuracy and long-term tracking

This article introduces SAM2Long, a novel approach to video object segmentation. By addressing the limitations of SAM2, SAM2Long utilizes a training-free memory tree structure to enhance long-term video segmentation, particularly in scenarios with occlusions and object re-appearances. This innovative method significantly improves the accuracy and robustness of video segmentation tasks.