TrackWhy / README.md
chinmay-1302's picture
Dataset card: link project page
3da4be3 verified
|
Raw History Blame Contribute Delete
4.14 kB
---
license: cc-by-nc-sa-3.0
task_categories:
- video-classification
tags:
- multi-object-tracking
- mot20
- failure-analysis
- vision-language-models
---
# TrackWhy: Tracker-Failure Clips
284 video clips of multi-object tracking failures, rendered from tracker output over
MOT20 sequences MOT20-01 and MOT20-03. Each clip centres on one failure event drawn from
a pool of 1,164 mined events, stratified at roughly 30 clips per condition category.
Companion release: annotations, attributor code and model outputs at
[https://github.com/chinmay-1302/trackwhy](https://github.com/chinmay-1302/trackwhy).
Paper: [https://openreview.net/pdf?id=TRJ7gFUnta](https://openreview.net/pdf?id=TRJ7gFUnta).
Project page: [https://chinmay-1302.github.io/trackwhy-page/](https://chinmay-1302.github.io/trackwhy-page/).
The full tree is about 3.4 GB.
## Contents
| Path | Description |
| -------------------- | --------------------------------------------------------------------------------- |
| `clips/` | Supervision-free rendering: highlight pinned to one tracker ID, never re-anchored |
| `clips_gt_informed/` | Ground-truth-informed rendering: highlight re-anchors to the matched track |
| `manifest.csv` | Clip ID, sequence, frame range, highlighted track IDs, rendering regime |
Each directory contains `C001.mp4`–`C284.mp4`. Clips span 81–201 frames (median 93),
play at 10 fps against a 25 fps source, and show tracker output only — every tracker
box with its numeric ID, the target highlighted. No ground-truth geometry appears in
any frame.
The Hub viewer’s `label` column is the source directory. The two regimes are paired
renderings of the same 284 clips, not classes; `manifest.csv` carries the real
per-clip metadata.
`manifest.csv` has one row per clip per regime (568 rows). For
`regime=supervision_free`, `highlight_after_track_id` is the successor identity that
the ground-truth-informed render re-anchors to; that highlight is not drawn in
`clips/`.
The two video directories both include `C205`. The companion ground-truth-informed
annotation reference omits that clip (283 records); the supervision-free reference
keeps it (284). The videos themselves are released for both regimes.
## The two rendering regimes
They differ at the failure and the difference is load-bearing. Under
ground-truth-informed rendering the highlight re-anchors to whichever tracker box the
ground truth now matches, so the visualisation appears continuous and the only cue
that anything happened is the change of label digits. Under supervision-free
rendering the highlight is pinned to one tracker identity and ends with it, and the
successor appears as an ordinary unhighlighted box among the others.
A reader who mixes the two directories will not reproduce the paper's numbers.
## Licence and attribution
These clips are derivative works of the MOT20 dataset and are released under
**CC BY-NC-SA 3.0**, the same licence MOT20 carries.
MOT20 is published by the Technical University of Munich Computer Vision Group under
CC BY-NC-SA 3.0. If you use this data, cite MOT20 as well as this work:
```bibtex
@article{dendorfer2020mot20,
title={MOT20: A benchmark for multi object tracking in crowded scenes},
author={Dendorfer, Patrick and Rezatofighi, Hamid and Milan, Anton and Shi, Javen
and Cremers, Daniel and Reid, Ian and Roth, Stefan and Schindler, Konrad
and Leal-Taix{\'e}, Laura},
journal={arXiv preprint},
year={2020}
}
```
```bibtex
@inproceedings{mhatre2026what,
title={What Counts as a Tracking Failure? Measuring Disagreement Between Metrics, Human Observers, and Vision-Language Models},
author={Chinmay Mhatre and Sankalp Gawali and Punit Rathore},
booktitle={Grounded and Faithful Vision-Language Models for Real-World Deployment 2026},
year={2026},
url={https://openreview.net/forum?id=TRJ7gFUnta}
}
```
Note: the MOTChallenge evaluation server is no longer in operation; dataset archives
remain available at their published addresses.