Utils/DATA_DOWNLOAD.md

# PhysInOne Rendered Data Download

## Choose your data

PhysInOne provides three separate public releases. Choose the guide that matches
the resources you need:

| Release | Contents | Download guide |
| --- | --- | --- |
| **1. Rendered Data** | Released Train and Val scene archives containing rendered observations and annotations | **This guide** |
| **2. 3D Assets** | Unreal Engine project files, asset libraries, Scenes, and Trajectories | [3D Assets: download and project setup](https://huggingface.co/datasets/vLAR/PhysInOne/blob/main/Utils/3D_ASSETS_DOWNLOAD.md) |
| **3. Leaderboard Evaluation Data** | Public inputs for Video Generation, Future Prediction, Physical Properties Estimation, and Motion Transfer; private ground truth is excluded | [Leaderboard: task inputs and download commands](https://huggingface.co/datasets/vLAR/PhysInOne/blob/main/Utils/LEADERBOARD_DOWNLOAD.md) |

The instructions below cover **Rendered Data only**. Use `filter_cases.py` to
select scenes, then `download_selected.py` to download their archives.
`download_data.py --task all` covers the separate 3D Assets and Leaderboard
releases; it does **not** download the full rendered dataset.

## Requirements and tools

- Python 3.9 or newer and Git for the setup command below.
- Enough disk space for the scene archives you select.
- No Python package installation, account, or authentication is required.

Clone the repository and run all commands from its root:

```bash
git clone https://github.com/vLAR-group/PhysInOne.git
cd PhysInOne
python scripts/filter_cases.py --help
python scripts/download_selected.py --help
```

The tools are also available under [Hugging Face `Utils/scripts/`](https://huggingface.co/datasets/vLAR/PhysInOne/tree/main/Utils/scripts).
Keep the script directory together: `download_selected.py` imports shared
functions from `download_data.py`, and both the scene index and repository map
are supplied under `scripts/download_lists/`. These are local dependencies;
the scripts do not install packages or fetch supporting scripts at runtime.

## Full rendered dataset selection

### Where the data is stored

| Split | Remote location | Selection |
| --- | --- | --- |
| Train | The 16 shard repositories listed on the [dataset homepage](https://huggingface.co/datasets/vLAR/PhysInOne#dataset-repositories) | `--split train` |
| Val | The main repository's [Val directory](https://huggingface.co/datasets/vLAR/PhysInOne/tree/main/Val) | `--split val` |

The bundled index lists the available scenes; the repository map resolves their
download locations automatically. Selection counts reflect that bundled index,
not a live scan of Hugging Face. Update your checkout when new data is released.
The released Val archive count is separate from the 103-scene Future Prediction
Leaderboard task.

### Example A: select and download a Train subset

This example selects up to 100 double-physics scenes containing `FrictionStop`.
The seed makes the selection reproducible for the same index.

```bash
python scripts/filter_cases.py \
  --split train \
  --activity_type double \
  --phenomena FrictionStop \
  --num 100 \
  --seed 42 \
  --output selected_train_cases.json \
  --show_stats
```

Preview the selection's download destinations:

```bash
python scripts/download_selected.py \
  --selection selected_train_cases.json \
  --output-dir ./PhysInOne_data/rendered_data \
  --keep-shard-structure \
  --dry-run
```

Download the selected archives:

```bash
python scripts/download_selected.py \
  --selection selected_train_cases.json \
  --output-dir ./PhysInOne_data/rendered_data \
  --keep-shard-structure
```

To select all indexed Train scenes, use `--split train` without
`--activity_type`, `--phenomena`, or `--num`. Preview the resulting selection
before starting a large download.

### Example B: select and download all indexed Val scenes

```bash
python scripts/filter_cases.py \
  --split val \
  --output selected_val_cases.json \
  --show_stats

python scripts/download_selected.py \
  --selection selected_val_cases.json \
  --output-dir ./PhysInOne_data/rendered_data \
  --keep-shard-structure
```

Use `val`, not `valid`. Add `--dry-run` to the download command to preview the
transfers. The selection command writes JSON only and does not download images.

## Selection options

These options belong to `filter_cases.py`:

| Option | Meaning |
| --- | --- |
| `--split train` / `--split val` | Select a rendered-data split. Omitting this option includes all indexed splits. |
| `--activity_type single` / `double` / `triple` | Filter by physical-interaction complexity. |
| `--phenomena NAME [NAME ...]` | Filter using phenomenon abbreviations, such as `FrictionStop`. |
| `--match_mode contains` | Require all requested phenomena; allow additional phenomena. This is the default. |
| `--match_mode exact` | Require exactly the requested set of phenomena. |
| `--num N` | Randomly select at most N matching scenes. Omit to keep all matches. |
| `--seed N` | Random sampling seed. Default: `42`. |
| `--output PATH` | Selection JSON destination. Default: `selected_cases.json`. |
| `--show_stats` | Print selected counts by split, complexity, and repository-map key. |
| `--assignment_file PATH` | Use a custom scene index. Default: `scripts/download_lists/repo_assignment.txt`. |

The parser also accepts `--split test`, but it only searches the scene index.
It is not a way to download the four Leaderboard task packages; use the
[Leaderboard guide](https://huggingface.co/datasets/vLAR/PhysInOne/blob/main/Utils/LEADERBOARD_DOWNLOAD.md) for those inputs.

## Download options

These options belong to `download_selected.py`:

| Option | Meaning |
| --- | --- |
| `--selection PATH` | Read the selection JSON. Default: `selected_cases.json`. |
| `--output-dir PATH` | Set a local output root without whitespace. The examples use `./PhysInOne_data/rendered_data`; the script default is `PhysInOne_data/full_dataset`. |
| `--keep-shard-structure` | Preserve repository-map key, split, and physics category directories. Without this option, ZIPs go directly into the output root. |
| `--workers N` | Concurrent downloads. Default: `4`. |
| `--retries N` | Retries after a failed request. Default: `3`. |
| `--timeout SECONDS` | Request timeout. Default: `60`. |
| `--revision REV` | Remote repository revision. Default: `main`. |
| `--no-resume` | Restart incomplete `.part` files instead of resuming. |
| `--overwrite` | Download again even when a completed local file exists. |
| `--dry-run` | Show the transfer count and up to 20 planned paths without downloading. |
| `--repo-map PATH` | Use a custom repository map. Default: `scripts/download_lists/repo_map.json`. |

## Downloaded files and logs

With `--keep-shard-structure`, the example commands produce this layout:

```text
PhysInOne_data/
└── rendered_data/
    ├── physinone_part1/
    │   └── Train/
    │       └── DoublePhysics/
    │           └── <scene>_trajectory.zip
    ├── ...
    ├── physinone_val/
    │   └── Val/
    │       └── <Physics>/
    │           └── <scene>_trajectory.zip
    └── _download_logs/
        └── YYYYMMDD_HHMMSS/
            ├── downloads.tsv
            ├── failures.tsv
            └── summary.json
```

Only directories needed by your selection are created. `physinone_val` is the
local repository-map key for Val data hosted in the main HF repository.

`download_selected.py` keeps the downloaded scene ZIPs compressed; it does not
automatically extract them. Extract the archives you need with a ZIP tool,
preserving their internal scene and camera directories. Its options differ from
the separate `download_data.py` used for 3D Assets and Leaderboard inputs.

Interactive terminals display file progress, transferred size, speed, and an
estimated remaining time. Redirected output uses periodic text updates. Use
`downloads.tsv`, `failures.tsv`, and `summary.json` to review individual files
and the overall result.

## Resume and troubleshooting

- Run the same download command again to reuse completed files and resume
  incomplete `.part` files. Files are renamed to their final names after download.
- If a completed ZIP needs replacement, use `--overwrite`.
- For a slow or rate-limited connection, lower concurrency with `--workers 2`.
- If a filter matches zero cases, check the spelling (`val`, not `valid`),
  update your checkout, and check that the requested scenes are included in
  `scripts/download_lists/repo_assignment.txt`.
- Keep `filter_cases.py`, `download_selected.py`, `download_data.py`, and
  `download_lists/` from the same release so that selection and repository
  mapping stay consistent.