Track Anything
Click-to-track video segmentation built on SAM and XMem, unmaintained since 2023 and replaced by SAM 2.
About
Within weeks of Meta's Segment Anything launch in April 2023, the SUSTech VIP Lab team behind Track Anything showed how to make it work on video: click an object in one frame, let SAM produce a mask, then hand that mask to the XMem tracker to follow it through the clip, with an optional E2FGVI pass to paint the object out. The Gradio app made interactive video masking and object removal reachable for people who had never trained a model, and it became a popular annotation shortcut. Code changes stopped in June 2023; the only later commit, in December 2025, added a README pointer to the authors' newer SAM-Body4D project. It can still be installed on an NVIDIA GPU, but its requirements are unpinned, so a fresh environment pulls current Gradio and PyTorch releases the 2023 code never saw, and nobody upstream is fixing the breakage. Its own code is MIT, while the E2FGVI inpainting model it relies on is CC BY-NC 4.0, so object removal is limited to non-commercial use. For new work, Meta's SAM 2 (Apache-2.0) handles click-to-track video segmentation in a single model, and SAM 3 adds text prompts.
Should you use Track Anything?
Pick it when
Pick Track Anything when you want a click-driven interface to mask and follow objects through a video for annotation, editing, or object removal, and prefer correcting results mid-clip over writing tracking code.
Look elsewhere when
Skip it for commercial object removal: the code is MIT, but its inpainting step uses E2FGVI under CC BY-NC 4.0. For pipelines, SAM 2 does video masks in one Apache 2.0 model, and Grounded SAM 2 suits text-driven tracking.
Alternatives to Track Anything
- Segment Anything 2
Video mask tracking built into the model under Apache 2.0, a safer base for products, though you build or adopt your own click interface.
- Grounded SAM 2
Selects objects from text rather than clicks and tracks them with SAM 2, better for batch jobs than interactive correction.
- CVAT
MIT-licensed multi-user annotation platform with review workflows and 20-plus export formats, heavier to host than Track Anything's standalone tool.
- CoTracker
Follows points rather than masks through occlusions, suited to motion analysis, but CC-BY-NC limits its main weights to non-commercial use.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- MIT
- Minimum VRAM
- 8 GB
- Added
- Apr 3, 2026