Multi Person Tracking In Group Footage

via “multi-object video segmentation with independent prompt-per-object tracking”

Meta's foundation model for visual segmentation.

Unique: Maintains independent memory buffers per tracked object, allowing the same cross-frame attention mechanism to operate on object-specific feature sequences. This design avoids global memory conflicts and enables flexible object-level prompting without requiring a unified object registry.

vs others: More flexible than traditional multi-object tracking (MOT) methods because it doesn't require pre-computed detections or appearance models; instead, it directly propagates semantic masks, handling appearance changes and occlusions through learned attention patterns.