README.md

---
language:
- en
tags:
- video
- egocentric
- world-model
- minecraft
size_categories:
- n<1K
---

# PWM-Bench

First-person source videos with aligned captions and action annotations for
scene-specific video model training and evaluation.

## Data

|Category|Content|Train videos|Test videos|
|---|---|---:|---:|
|Indoor|Real indoor first-person recordings|50|50|
|Outdoor|Real outdoor first-person recordings|50|50|
|Gaming|Minecraft/MineDojo-rendered first-person recordings|50|50|
|**Total**||**150**|**150**|

**300 videos, 33,750 frames.** All videos are 832×480, 16 fps, H.264 MP4,
with no audio. Each scene has train80 (5 seconds) and test145 (9.0625 seconds).
There is **no validation split**.

Indoor/Outdoor clips are segmented from existing real egocentric recordings
(EgoVid and the extended source pool). Gaming clips are captured from rendered
Minecraft/MineDojo environments with simulator poses. All are organized into
train/test pairs and paired with window-level captions. Movement and camera
tokens are extracted from those captions. Source frames are retained in H5 and
encoded as MP4 for viewing; no interpolation or padding is applied.

These are within-scene train/test pairs. The collection is curated, and different
scenes may share an upstream source or game location; source information is in
`manifest.json`.

## Files

```text
README.md
manifest.json             # scene list, format and source information
splits.json               # train/test video paths
SHA256SUMS                # file integrity checks
indoor/indoor1/
  train.mp4
  test.mp4
  data.h5
  annotations.json
  train_caption.txt
  test_caption.txt
indoor/indoor2/ ... indoor50/
outdoor/outdoor1/ ... outdoor50/
gaming/gaming1/ ... gaming50/
```

Each TXT is UTF-8, with one complete caption per line and no header: six lines
for train, five for test. TXT, JSON and H5 captions match exactly.

## Annotations

|Split|Clip frames|Windows per scene|Window length|Start frames|
|---|---:|---:|---:|---|
|train|80|6|29|0, 9, 18, 27, 36, 45|
|test|145|5|29|0, 29, 58, 87, 116|

Indices are zero-based and clip-local; ends are exclusive. Train windows overlap;
the final window is `[45:74]`, leaving six source frames without an additional
annotated window. Test windows cover all 145 frames. Use the explicit bounds,
not a per-frame or implicit 9-frame caption index.

H5 stores `<scene_id>/<train|test>/`, where the scene ID matches its folder:

- `video_clip`: JPEG byte arrays, 80 or 145 entries; decode to RGB for training.
- `chunk_captions`, `chunk_keys`, `chunk_mouse`: six train or five test entries.
- `chunk_start_frames`, `chunk_end_frames_exclusive`: window boundaries.
- `poses`: source matrices, shaped `(N,4,4)`. **Real-video poses are placeholders,
  not ground-truth trajectories**; Gaming uses simulator matrices.
- `prompt`: auxiliary scene description. Do not prepend it to captions, which
  already contain the full conditioning text.

`annotations.json` contains `scene_id`, `train_windows` and `test_windows`.
Each window has caption, key, camera, frame bounds and `motion_scalars`
(distance, turn-speed, view-rotation-speed).

Movement tokens: W/S/A/D = forward/backward/left/right; combinations use `+`.
`·` means stationary. Camera tokens: `·` = no turn, `←` = left, `→` = right.
These are conditioning labels; motion numbers and speed-unit wording in captions
are not calibrated physical speeds.

## Loading

Requires Python, h5py, NumPy and OpenCV. Run from the dataset directory.

```python
from pathlib import Path
import h5py
import cv2
import numpy as np

def load_windows(folder, split="train"):
    folder = Path(folder)
    with h5py.File(folder / "data.h5", "r") as f:
        g = f[f"{folder.name}/{split}"]
        for i, caption in enumerate(g["chunk_captions"].asstr()[:]):
            start = int(g["chunk_start_frames"][i])
            end = int(g["chunk_end_frames_exclusive"][i])
            video = np.stack([
                cv2.cvtColor(cv2.imdecode(jpeg, cv2.IMREAD_COLOR),
                             cv2.COLOR_BGR2RGB)
                for jpeg in g["video_clip"][start:end]
            ])
            assert video.shape == (29, 480, 832, 3)  # uint8 RGB
            yield video, caption

for video, caption in load_windows("indoor/indoor1", "train"):
    pass  # Apply your model's normalization and train with this pair.
# Use split="test" for the five test windows.
```

## Sources and rights

Please respect the rights and applicable terms of use of the original content and assets.