Meta releases SAM 3.1 for faster real-time video tracking
Meta announced SAM 3.1 — an updated version of the Segment Anything Model (SAM 3), designed to significantly improve the efficiency and accessibility of video processing.
This new iteration introduces object multiplexing — a key innovation that allows the model to track up to 16 objects in a single pass. This architectural change eliminates redundant computations and memory bottlenecks, leading to a substantial increase in throughput: specifically, doubling the processing speed for videos with average objects from 16 to 32 frames per second on an H100 GPU.
SAM 3.1 provides real-time object tracking in complex, crowded videos while simultaneously reducing overall GPU resource requirements. This makes high-performance AI applications more feasible on more compact and accessible hardware. Furthermore, Meta continues to expand the SAM ecosystem by releasing SAM 3D — a toolkit for 3D reconstruction of objects and people, and launching the Segment Anything Playground for broader community experimentation.
Why it matters
- —Significantly increases video processing efficiency, doubling throughput.
- —Implementing object multiplexing makes high-performance AI accessible on less powerful hardware.
- —Advances the entire SAM ecosystem with 3D reconstruction capabilities.
Key facts
- SAM 3.1 is a replacement for SAM 3.
- Uses object multiplexing to track up to 16 objects in a single pass.
- Processing speed increases from 16 to 32 FPS on H100 GPU.
- SAM 3.1 provides real-time object tracking in complex scenes.
- Meta also released SAM 3D for 3D reconstruction.
The full text is in the original source. Here we provide a brief summary and key facts.
