CoTracker
MetaCoTracker is Meta AI's transformer model for tracking points through video: given any pixels in a frame, picked by hand or sampled on a grid, it follows them jointly across the clip. CoTracker3, the current version, comes in online and offline variants and was trained on synthetic Kubric video, then fine-tuned on real videos with pseudo-labels. It loads through torch.hub.
Openness
2 high confidence- weights
- open (CoTracker3 baseline and scaled, online and offline checkpoints on the Hub, ungated)
- data
- documented-not-released (the synthetic Kubric training set is on the Hub; the real videos used for pseudo-label fine-tuning are not released)
- code
- open (training scripts for Kubric and for pseudo-label fine-tuning)
- license
- CC-BY-NC-4.0 (most of the code and the checkpoints; a few third-party parts are MIT or Apache-2.0)
CoTracker's checkpoints and most of its code are under CC BY-NC 4.0, which forbids commercial use. Meta publishes the training scripts and its synthetic Kubric training set, but not the real videos the released models were fine-tuned on.
- https://huggingface.co/api/datasets/facebook/CoTracker3_Kubric recorded 2026-11-12
Hub record for facebook/CoTracker3_Kubric: public, not gated, license apache-2.0; "This dataset was specifically created for training CoTracker 3".
- https://huggingface.co/api/models/facebook/cotracker3 recorded 2026-11-12
Hub record for facebook/cotracker3: gated false, license cc-by-nc-4.0, baseline and scaled online and offline .pth checkpoints.
- https://raw.githubusercontent.com/facebookresearch/co-tracker/main/LICENSE.md recorded 2026-11-12
LICENSE.md is Creative Commons Attribution-NonCommercial 4.0 International.
- https://raw.githubusercontent.com/facebookresearch/co-tracker/main/README.md recorded 2026-11-12
README: "The majority of CoTracker is licensed under CC-BY-NC, however portions of the project are available under separate license terms"; training commands for train_on_kubric.py; "In order to launch training with pseudo-labelling, you need to collect your own dataset of real videos."
- https://ungh.cc/repos/facebookresearch/co-tracker/files/main recorded 2026-11-12
Repository tree: train_on_kubric.py, train_on_real_data.py and the four launch_training_*.sh scripts.
Adoption
2 high confidenceHugging Face downloads over the trailing 30 days for the CoTracker3 and earlier CoTracker repositories, almost all of them for CoTracker3. Loading through torch.hub fetches the same Hub files.
- https://huggingface.co/api/models?author=facebook&search=cotracker&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
facebook/cotracker3 44,095 and facebook/cotracker 1,636 downloads in the trailing 30 days (45,731 together).
Capability
4 medium confidenceCoTracker follows whichever points the user picks in their own video, with no training step, much as Segment Anything segments whatever the user points at. It does point tracking only, where Ultralytics spans detection, segmentation and depth.
- https://huggingface.co/facebook/cotracker3/raw/main/README.md recorded 2026-11-12
cotracker3 card: "CoTracker can track: Any pixel in a video; A quasi-dense set of pixels together; Points can be manually selected or sampled on a grid in any video frame".
- https://raw.githubusercontent.com/facebookresearch/co-tracker/main/README.md recorded 2026-11-12
README results table: CoTracker3 offline 76.9 and online 76.7 on DAVIS, 67.8 and 68.3 on Kinetics, with the released checkpoints.
Verified 2026-11-12