Frame By Frame Pose Tracking With Temporal Keypoint Output

via “streaming memory-augmented video object tracking across frames”

Meta's foundation model for visual segmentation.

Unique: Uses a streaming memory architecture where frame features are compressed and stored in a fixed-size buffer, with cross-frame attention enabling mask propagation without re-encoding. This design treats video as a sequence of single-frame images processed through a unified architecture, avoiding separate video-specific models.

vs others: More efficient than optical flow-based tracking (e.g., DeepFlow) because it directly propagates semantic masks through learned attention rather than computing pixel-level motion, reducing computational overhead while maintaining temporal consistency across diverse object types.