---
configs:
  - config_name: track
    default: true
    data_files:
      - split: train
        path: track/train-*
license: apache-2.0
task_categories:
  - video-classification
  - object-detection
tags:
  - video-object-tracking
  - video-segmentation
  - synthetic
---

# MolmoPoint-TrackSyn Dataset

Synthetic point tracking annotations for procedurally generated videos generated with Blender.

Each example contains an expression describing an object, per-frame point trajectories, and video metadata. All videos are encoded as **6 FPS** and points are sampled at **2 FPS**.

## Dataset Statistics

| Video Source | Unique Annotations | Unique Videos |
|-------------|-------------------|-----------------|
| static-camera | 34,324            | 11,629        |
| dyna-camera   | 41,841            | 14,158        |
| **Total**     | **76,165**        | **25,787**    |

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `id` | `string` | Unique example identifier |
| `video` | `string` | Relative video path (without extension), e.g. `static-camera/{run_dir}/{video_file}`. We support static camera (`static-camera`) and dynamic camera (`dyna-camera`) setups. |
| `expression` | `string` | Natural-language description of the tracked object |
| `fps` | `int64` | Original video FPS |
| `sampling_fps` | `int64` | Sampling FPS used for annotation (always 2) |
| `height` | `int64` | Video height in pixels |
| `width` | `int64` | Video width in pixels |
| `n_frames` | `int64` | Number of frames in the sampled clip |
| `task` | `string` | Task type (always `"track"`) |
| `frame_trajectories` | `list[object]` | Per-frame point tracks (frame index, timestamp, point coords + occlusion) |
| `mask_id` | `list[string]` | Optional mask identifiers |
| `obj_id` | `list[int64]` | Optional object identifiers |

## Video Download

Videos are bundled in this repository as `synthetic_tracks.tar`.

### Automatic download

```python
from olmo.data.molmo2_video_track_datasets import MolmoPointTrackSyn

# Downloads the tar from HF, extracts, and verifies
MolmoPointTrackSyn.download()
```

### Manual download

```bash
# Download the tar from HuggingFace
huggingface-cli download allenai/MolmoPoint-TrackSyn synthetic_tracks.tar --repo-type dataset --local-dir ./MolmoPoint-TrackSyn

# Extract
tar -xf ./MolmoPoint-TrackSyn/synthetic_tracks.tar -C ./MolmoPoint-TrackSyn/
```

After extraction the directory structure is:

```
MolmoPoint-TrackSyn/
├── static-camera/
│   ├── {run_dir}/
│   │   ├── video.mp4
│   │   └── metadata.json
│   └── ...
└── dyna-camera/
    ├── {run_dir}/
    │   ├── video.mp4
    │   └── metadata.json
    └── ...
```

The `video` column maps directly to the file path: `{VIDEO_HOME}/{video}/video.mp4

## Usage

```python
from datasets import load_dataset

# Load the dataset
ds = load_dataset("allenai/MolmoPoint-TrackSyn", split="train")

# Inspect an example
print(ds[0])
```

## Citation

If you use this dataset, please cite the MolmoPoint paper.

## License
Dataset license: ODC-BY
Dataset card (License section): This dataset is licensed under ODC-BY. 
It is intended for research and educational use in accordance with [Ai2’s Responsible Use Guidelines](https://allenai.org/responsible-use).