# Track3D sources and attribution

This page presents saved, locally computed **TraceAnything** experiments on DAVIS video sequences. It is not a TrackEverything implementation, an official model demo, or a new-video inference service.

## Video and foreground annotations

**DAVIS 2016** — Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. *A Benchmark Dataset and Evaluation Methodology for Video Object Segmentation*, CVPR 2016.

- [Official DAVIS download and citation page](https://davischallenge.org/davis2016/code.html)
- [Official dataset archive](https://graphics.ethz.ch/Downloads/Data/Davis/DAVIS-data.zip)
- [CC BY-NC 4.0 license](https://creativecommons.org/licenses/by-nc/4.0/), linked by the archive README's Terms of Use.
- Sequences shown: `train` (80 frames) and `dance-twirl` (90 frames). The original RGB frames and official binary foreground annotations came from the ETH DAVIS host. The original frames are encoded as derived MP4s and resized for inference and display. `dance_crop` additionally applies a fixed crop around the person; `dance_sampled` and `dance_crop` retain every third source frame.
- The chosen preview rate is 24 fps, or 8 fps for every-third-frame variants. This is not a verified original capture rate.
- The public overlay videos, independent-viewpoint renders, thumbnails, and 3D visualizations are derived experiment outputs with color points, trajectories, and camera projection added by this lab. Foreground annotations select displayed tracks and weight camera fitting; they also determine the fixed person crop. They do not provide 3D ground truth or correct the predicted coordinates.

The archive README's Credits paragraph says Attribution 4.0, while its Terms of Use links BY-NC 4.0. This research demonstration retains attribution and the non-commercial condition from Terms of Use. `train` and `dance-twirl` are not among the README's separately credited YouTube sequences.

## Tracking model

**Trace Anything: Representing Any Video in 4D via Trajectory Fields** — Xinhang Liu, Yuxi Xiao, Donny Y. Chen, Jiashi Feng, Yu-Wing Tai, Chi-Keung Tang, and Bingyi Kang (2025).

- [Official source repository](https://github.com/ByteDance-Seed/TraceAnything)
- [Paper](https://arxiv.org/abs/2510.13802)
- [Official weights](https://huggingface.co/depth-anything/trace-anything)
- Source revision: `54677b5e7bf11510c2e8c917a509988ad379f8eb`.
- Source code: [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).
- Model weights: [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). No model weights are included in this static website.

Inference uses the official weights with a Windows-compatible PyTorch attention backend and memory-saving execution scheduling. The viewer retains the predicted 3D coordinates; supplied annotations and camera fitting only affect the declared crop, track selection, and display. Predictions have arbitrary scale, uncertain visibility, and visible tracking errors, particularly during dance rotations and occlusion. These scenes do not establish 999-frame tracking or ground-truth accuracy.

## Public result metadata

Each scene includes `scene.json`, `verification.json`, and camera-alignment metadata under `overlay/`. Source frame indices, media hashes, measurements, coordinate files, and export checks are retained. Provenance paths such as `runs/...` and `inputs/source_media/...` are relative labels describing the original experiment tree; they are not download URLs. Machine-specific absolute paths were removed for publication. Full original records remain in the local experiment archive.
