Utils/LEADERBOARD_DOWNLOAD.md

# PhysInOne Leaderboard Evaluation Data

This guide explains how to download the public evaluation data for all four
PhysInOne Leaderboard tasks:

- Video Generation
- Future Prediction
- Physical Properties Estimation
- Motion Transfer

The data is hosted in the public
[`vLAR/PhysInOne`](https://huggingface.co/datasets/vLAR/PhysInOne) dataset
repository under [`Leaderboard/`](https://huggingface.co/datasets/vLAR/PhysInOne/tree/main/Leaderboard).
No Hugging Face access token is required for public downloads.

## Install

Python 3.9 or newer is recommended. The downloader uses only the Python
standard library; no third-party package or Hugging Face authentication is
required.

Clone the GitHub repository, which contains the downloader and all maintained
manifests:

```bash
git clone https://github.com/vLAR-group/PhysInOne.git
cd PhysInOne
```

The same script tree is mirrored under `Utils/scripts/` in the Hugging Face
dataset repository.

## Progress display

Interactive terminals show a single live progress bar with completed file count,
transferred and estimated total size, current download speed, and ETA. Because
the manifests do not store object sizes, the total size and ETA are estimates
that become more accurate as Hugging Face responses provide file sizes. Redirected
output automatically uses periodic plain-text snapshots, and complete per-file
records remain available under `_download_logs/`.

## Quick start

Run these commands from the cloned GitHub repository root.

Download all four tasks:

```bash
python scripts/download_data.py \
  --task all \
  --output-dir ./PhysInOne-Leaderboard
```

Download one task:

```bash
python scripts/download_data.py \
  --task future-prediction \
  --output-dir ./PhysInOne-Leaderboard
```

Download multiple tasks in one run:

```bash
python scripts/download_data.py \
  --task video-generation motion-transfer \
  --output-dir ./PhysInOne-Leaderboard
```

## Task names and file lists

For each selected task, the downloader reads a maintained TXT manifest bundled
under `scripts/download_lists/`. It does not fetch utility files at runtime and
does not scan unrelated parts of the large PhysInOne repository. The Hugging
Face copies below are provided for inspection and reproducibility.

| `--task` value | Repository directory | Maintained file list |
| --- | --- | --- |
| `video-generation` | `Leaderboard/Video Generation/` | [`video-generation.txt`](https://huggingface.co/datasets/vLAR/PhysInOne/blob/main/Utils/leaderboard_lists/video-generation.txt) |
| `future-prediction` | `Leaderboard/Future Prediction/` | [`future-prediction.txt`](https://huggingface.co/datasets/vLAR/PhysInOne/blob/main/Utils/leaderboard_lists/future-prediction.txt) |
| `physical-properties-estimation` | `Leaderboard/Physical Properties Estimation/` | [`physical-properties-estimation.txt`](https://huggingface.co/datasets/vLAR/PhysInOne/blob/main/Utils/leaderboard_lists/physical-properties-estimation.txt) |
| `motion-transfer` | `Leaderboard/Motion Transfer/` | [`motion-transfer.txt`](https://huggingface.co/datasets/vLAR/PhysInOne/blob/main/Utils/leaderboard_lists/motion-transfer.txt) |

### Copyable task commands

| Task | Command | Space-free local directory |
| --- | --- | --- |
| Video Generation | `python scripts/download_data.py --task video-generation --output-dir ./PhysInOne-Leaderboard` | `PhysInOne-Leaderboard/leaderboard/video_generation/` |
| Future Prediction | `python scripts/download_data.py --task future-prediction --output-dir ./PhysInOne-Leaderboard` | `PhysInOne-Leaderboard/leaderboard/future_prediction/` |
| Physical Properties Estimation | `python scripts/download_data.py --task physical-properties-estimation --output-dir ./PhysInOne-Leaderboard` | `PhysInOne-Leaderboard/leaderboard/physical_properties_estimation/` |
| Motion Transfer | `python scripts/download_data.py --task motion-transfer --output-dir ./PhysInOne-Leaderboard` | `PhysInOne-Leaderboard/leaderboard/motion_transfer/` |
| All four tasks | `python scripts/download_data.py --task all --output-dir ./PhysInOne-Leaderboard` | `PhysInOne-Leaderboard/leaderboard/` |

Use `--list-only` to print the selected public URLs without downloading data:

```bash
python scripts/download_data.py \
  --task motion-transfer \
  --list-only
```

Use `--dry-run` to show the number of selected files:

```bash
python scripts/download_data.py \
  --task all \
  --dry-run
```

## Download individual scenes

`--scene` performs a case-insensitive substring match against repository paths.
It accepts a complete scene name or the unique six-character scene ID and can
be repeated.

```bash
python scripts/download_data.py \
  --task future-prediction \
  --scene yUlWPB \
  --output-dir ./PhysInOne-Leaderboard
```

```bash
python scripts/download_data.py \
  --task motion-transfer \
  --scene B15ta6 \
  --scene 7omQio \
  --output-dir ./PhysInOne-Leaderboard
```

## ZIP extraction

By default, archives remain compressed. This saves disk space and preserves the
exact released artifacts. Add `--extract` to validate each ZIP with a CRC check
and safely extract it next to the archive:

```bash
python scripts/download_data.py \
  --task physical-properties-estimation \
  --output-dir ./PhysInOne-Leaderboard \
  --extract
```

To remove an archive after successful extraction, explicitly add
`--delete-zip-after-extract`:

```bash
python scripts/download_data.py \
  --task future-prediction \
  --output-dir ./PhysInOne-Leaderboard \
  --extract \
  --delete-zip-after-extract
```

Allow enough free space before extracting. Keeping both a ZIP and its extracted
contents can temporarily require approximately twice the archive's disk space.

## Released contents

### Video Generation

Each scene directory contains the first RGB frame from the selected static
camera, the first RGB frame from the moving camera, the two corresponding
Blender camera JSON files, and `caption.txt`.

### Future Prediction

Each scene ZIP contains only fixed training views. It includes RGB frames
`0000.jpg` through `0074.jpg` and the three split descriptions
`transforms_train.json`, `transforms_val.json`, and `transforms_test.json`.
Novel views and `CineCamera_Moving` are excluded from these public inputs.

### Physical Properties Estimation

Each scene ZIP contains background images for novel views; background, RGB, and
segmentation data for training views; camera JSON files for cameras 0 through
14; `particle_materials.json`; and `recorder_stats.json`. Novel views are
selected from each scene's transform metadata rather than assumed to have fixed
camera indices. Shared supplementary resources are stored under
`test_supp_info/`.

### Motion Transfer

Each scene ZIP contains source-video RGB frames, the first reference frame for
the target/origin scene, and `caption.txt`, organized by the two selected camera
views.

## Output layout

The downloader keeps every scene-relative path unchanged but maps the four
remote task roots to stable, space-free local names:

```text
PhysInOne-Leaderboard/
|-- leaderboard/
|   |-- video_generation/
|   |-- future_prediction/
|   |-- physical_properties_estimation/
|   +-- motion_transfer/
+-- _download_logs/
    +-- YYYYMMDD_HHMMSS/
        |-- downloads.tsv
        |-- extractions.tsv
        |-- failures.tsv
        +-- summary.json
```

## Resume and troubleshooting

Downloads use resumable anonymous HTTPS requests. Re-running the same command
reuses completed files and resumes valid `.part` files. Review `failures.tsv`
after each run and repeat the same command to retry failed files.

The default is four concurrent downloads. For a slow or rate-limited network,
reduce concurrency:

```bash
python scripts/download_data.py \
  --task all \
  --output-dir ./PhysInOne-Leaderboard \
  --workers 2
```

Public downloads are anonymous. The downloader has no token parameter or
credential-handling path.