README.md

---
pretty_name: RoboTrack Real v3
tags:
  - video
  - robotics
  - computer-vision
  - point-tracking
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - train/metadata.parquet
          - train/*/video.mp4
          - train/*/point_track_vis.mp4
    drop_labels: true
---

# RoboTrack Real v3

RoboTrack Real v3 is an evaluation dataset of 318 real-world video clips with
sparse 2D point trajectories and visibility annotations. Each example includes
the original RGB video, a rendered visualization of its point tracks, and the
underlying NumPy annotation archive.

## Dataset Viewer

The viewer exposes two playable video columns:

- `raw`: the original RGB clip (`raw_file_name` in `metadata.parquet`)
- `visualization`: the RGB clip with colored, numbered tracks and one second of
  visibility-aware trajectory history (`visualization_file_name`)

The remaining columns provide clip dimensions, timing, annotation statistics,
and optional `review_status` and `review_notes` fields for annotator review.

## Layout

```text
train/
  metadata.parquet
  <clip_id>/
    video.mp4
    point_track_vis.mp4
    point_tracks.npz
scripts/
  visualize_robotrack_dataset.py
```

All media files are stored directly in this repository; there are no symbolic
links. Videos are H.264 with matching frame counts and timing between the raw
and visualization versions.

## Annotation format

Each `point_tracks.npz` contains:

- `trajs_2d`: `float32` array shaped `(T, N, 2)` containing pixel coordinates
  in `(x, y)` order
- `visibility`: `float32` array shaped `(T, N)`, where values greater than `0.5`
  are visible
- `query_frames`: `int32` array shaped `(N,)` containing the query frame for
  each track

Here, `T` is the number of video frames and `N` is the number of annotated
tracks. Invisible coordinates are stored as `(0, 0)`.

## Frame rates

- 174 clips at 15 FPS
- 1 clip at 20 FPS
- 143 clips at 30 FPS

Forty-five clips whose containers incorrectly reported 60 FPS were retimed to
15 FPS without dropping frames. Their H.264 streams were copied without lossy
re-encoding, and the visualization videos use the same corrected timing.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("<namespace>/robotrack-real-v3", split="train")
example = dataset[0]
print(example["clip_id"], example["num_tracks"])
```

The NPZ path for each example is available in `annotation_path`. The included
renderer can recreate the point-track videos if needed.