hn.today

Show HN: Open-source, real-time object tracking for live video feeds

github.com36 points2 comments
Screenshot of Show HN: Open-source, real-time object tracking for live video feeds

tamserve is an open-source serving engine for Meta’s SAM 2, SAM 2.1, and EdgeTAM video-tracking models. It adapts models designed to process one recorded video at a time for continuous, multi-camera feeds: a single model instance batches work across cameras and objects, while retaining tracking state between frames. Applications provide the latest decoded frame from each camera and receive pixel-level object masks; they remain responsible for reading the camera streams. Objects can be started or stopped at any time using a box, clicks, or an existing mask. A separate Tracker can identify objects from text descriptions and follow matches across cameras, and a router can distribute camera workloads across multiple GPUs.

On an RTX 4090, the project reports that EdgeTAM uses 2.2 GB of GPU memory for eight cameras versus 8.6 GB with the model’s original code, and processes 323 versus 280 total frames per second. For 64 cameras, it reports 321 frames per second in 5.7 GB; memory for an 800-frame video is 0.28 GB rather than 9.8 GB. Accuracy on DAVIS 2017 is nearly unchanged: 0.8685 versus 0.8686. A live-loop benchmark with eight 720p cameras reports 273 frames per second, or about 34 per camera. The package installs with pip, requires Python 3.10 or newer and an NVIDIA GPU for live performance, and downloads its model on first use.

Read on github.com2 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in Programming

Solid 2.0

Solid 2.0

Solid 2.0 is now stable, five years after version 1.0, with a smaller API and diagnostics designed for AI agents. Its reactive graph now handles asynchronous computations, while built-in server functions support RPC calls and persistent Server-Sent Events connections. (solidjs.com)

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.