Meta SAM 3D

Meta’s Segment Anything Model 2 (SAM 2) delivers real-time, unified mask generation for 3D spatio-temporal data and video point clouds.

SAM 2 expands the original Segment Anything capabilities into the third dimension by treating video as a continuous 3D spatio-temporal entity. It utilizes a streaming memory architecture (storing feature memories and object pointers) to track masks across frames with high precision. In benchmark tests, SAM 2 outperforms previous models using 3x less interaction time while maintaining 80+ frames per second performance. This technology enables developers to segment 3D volumes in medical imaging or track dynamic objects in robotics via simple point, box, or mask prompts.

https://github.com/facebookresearch/segment-anything-2