🤖 AI Summary
A new open-source tool, tamserve, has been announced, enabling real-time multi-camera object tracking for live video feeds using Meta’s models SAM 2 and EdgeTAM. Unlike previous implementations that required separate model instances for each camera—resulting in high memory usage and slower speeds—tamserve allows all cameras to share a single model instance. This innovative serving engine efficiently tracks multiple objects in real-time, significantly reducing GPU memory requirements and enhancing processing speed. For example, tracking across eight cameras now only demands 2.2 GB of GPU memory compared to 8.6 GB with traditional systems.
This development holds significant implications for various industries needing precise object tracking from multiple camera angles, such as sports analytics, retail monitoring, and robotics. The ability to dynamically track new objects mid-stream and access them via simple prompts (like pointing or naming) makes the technology more user-friendly. Tamserve’s architecture also optimizes GPU usage, achieving impressive performance metrics—tracking 321 frames per second across 64 cameras on an NVIDIA RTX 4090, while maintaining the same level of accuracy. Such advancements stand to enhance live video analytics, making it more accessible and efficient for diverse applications.
Loading comments...
login to comment
loading comments...
no comments yet