Fix validation data selection and downloads

Utils/DATA_DOWNLOAD.md

@@ -0,0 +1,301 @@
+# PhysInOne Public Data Download
+
+The PhysInOne public downloader retrieves the released Leaderboard inputs and
+released 3D assets from the public `vLAR/PhysInOne` dataset repository. It
+uses only the Python standard library and anonymous HTTPS requests.
+
+Ground-truth outputs for the Leaderboard are private and are not included.
+
+## Requirements
+
+- Python 3.9 or newer
+- Enough free disk space for the selected release
+- Unreal Engine 5.5.4 when using the 3D asset project
+
+No Python package installation or account configuration is required.
+
+## Get the downloader
+
+Clone the GitHub repository and run the script from the project root:
+
+```bash
+git clone https://github.com/vLAR-group/PhysInOne.git
+cd PhysInOne
+python scripts/download_data.py --help
+```
+
+The maintained download lists are stored in `scripts/download_lists/`. The
+script reads them locally and does not fetch utility files at runtime. A
+byte-identical public mirror of the complete script tree is maintained under
+[`Utils/scripts/`](https://huggingface.co/datasets/vLAR/PhysInOne/tree/main/Utils/scripts).
+
+## Progress display
+
+Interactive terminals show a single live progress bar with completed file count,
+transferred and estimated total size, current download speed, and ETA. The total
+size and ETA are marked as estimates because public manifests do not store object
+sizes; they become more accurate as Hugging Face responses provide file sizes.
+When output is redirected to a log, the downloader automatically writes periodic
+plain-text snapshots instead of terminal control characters. Per-file success and
+failure details are always retained under `_download_logs/`.
+
+## Quick start
+
+Download all four public Leaderboard tasks:
+
+```bash
+python scripts/download_data.py \
+  --task video_generation future_prediction \
+         physical_properties_estimation motion_transfer \
+  --output-dir ./PhysInOne_data
+```
+
+Download one task:
+
+```bash
+python scripts/download_data.py \
+  --task future_prediction \
+  --output-dir ./PhysInOne_data
+```
+
+Download and assemble all currently released 3D assets:
+
+```bash
+python scripts/download_data.py \
+  --task 3d_assets \
+  --output-dir ./PhysInOne_data
+```
+
+ZIP extraction is enabled by default for `3d_assets`. Leaderboard ZIP files
+remain compressed unless `--extract` is specified.
+
+## Task values
+
+| `--task` value | Released public content | File count |
+| --- | --- | ---: |
+| `video_generation` | Initial frames, camera metadata, and captions | 75,865 |
+| `future_prediction` | Fixed-view input sequences and split metadata | 103 ZIP files |
+| `physical_properties_estimation` | Scene inputs and shared supplementary files | 72 scene ZIP files plus 2 shared files |
+| `motion_transfer` | Source videos, reference frames, and captions | 214 ZIP files |
+| `3d_assets` | Unreal Engine project assets for 14,847 Val and 120,660 Train scenes | 4,585 files |
+| `all` | Every item above | 80,841 files |
+
+Multiple task values can follow one `--task` option. Hyphenated names and the
+short aliases shown by `--help` are also accepted.
+
+## Command-line options
+
+| Option | Description |
+| --- | --- |
+| `--task NAME [NAME ...]` | Select one or more tasks, or `all`. |
+| `--output-dir PATH` | Set the local output root. The path must not contain whitespace. |
+| `--scene TEXT` | Keep paths containing a scene name or six-character scene ID. Repeatable. |
+| `--workers N` | Set concurrent downloads. Default: `4`. |
+| `--revision REV` | Select a dataset revision. Default: `main`. |
+| `--retries N` | Set retries after each failed request. Default: `3`. |
+| `--timeout SECONDS` | Set the timeout for each network request. Default: `60`. |
+| `--extract` | Safely validate and extract selected ZIP files. |
+| `--no-extract` | Keep all ZIP files compressed, including 3D asset archives. |
+| `--delete-zip-after-extract` | Remove a ZIP only after successful extraction. |
+| `--no-resume` | Restart incomplete `.part` files instead of resuming. |
+| `--force` | Replace completed local downloads. |
+| `--list-only` | Print the selected public URLs and exit. |
+| `--dry-run` | Show counts and output locations without downloading. |
+
+## Examples
+
+Download Video Generation inputs:
+
+```bash
+python scripts/download_data.py \
+  --task video_generation \
+  --output-dir ./PhysInOne_data
+```
+
+Download and extract Future Prediction inputs:
+
+```bash
+python scripts/download_data.py \
+  --task future_prediction \
+  --extract \
+  --output-dir ./PhysInOne_data
+```
+
+Download selected Motion Transfer scenes by ID:
+
+```bash
+python scripts/download_data.py \
+  --task motion_transfer \
+  --scene B15ta6 \
+  --scene 7omQio \
+  --output-dir ./PhysInOne_data
+```
+
+Preview Physical Properties Estimation without downloading:
+
+```bash
+python scripts/download_data.py \
+  --task physical_properties_estimation \
+  --dry-run \
+  --output-dir ./PhysInOne_data
+```
+
+Download 3D assets and remove archives after verified extraction:
+
+```bash
+python scripts/download_data.py \
+  --task 3d_assets \
+  --delete-zip-after-extract \
+  --output-dir ./PhysInOne_data
+```
+
+## Full rendered dataset selection
+
+The repository also includes a filter and downloader for selecting scene
+archives from the public rendered dataset. Train archives are distributed
+across 16 shard repositories, while the currently released Val archives are
+stored in the main `vLAR/PhysInOne` dataset repository. The bundled index and
+repository map handle this difference automatically.
+
+Create a selection file:
+
+```bash
+python scripts/filter_cases.py \
+  --split train \
+  --activity_type double \
+  --phenomena FrictionStop \
+  --num 100 \
+  --output selected_cases.json
+```
+
+Select all currently released Val archives:
+
+```bash
+python scripts/filter_cases.py \
+  --split val \
+  --output selected_val_cases.json \
+  --show_stats
+```
+
+This currently selects 199 Val scene archives. The count follows the public
+files present under `Val/` and may increase as additional validation data is
+released.
+
+Download the selected archives:
+
+```bash
+python scripts/download_selected.py \
+  --selection selected_cases.json \
+  --output-dir ./PhysInOne_data/full_dataset \
+  --keep-shard-structure
+```
+
+`filter_cases.py` reads the bundled
+`scripts/download_lists/repo_assignment.txt` index. `download_selected.py`
+reads the bundled `repo_map.json` and uses the same anonymous, resumable HTTPS
+implementation as the task downloader.
+
+| Selection option | Description |
+| --- | --- |
+| `--split train|val|test` | Filter by dataset split. |
+| `--activity_type single|double|triple` | Filter by interaction complexity. |
+| `--phenomena NAME [NAME ...]` | Require one or more physical phenomena. |
+| `--match_mode contains|exact` | Control phenomenon matching. |
+| `--num N` | Randomly sample at most `N` matching cases. |
+| `--seed N` | Set the sampling seed. |
+| `--output PATH` | Write the selection JSON. |
+
+| Selected-download option | Description |
+| --- | --- |
+| `--selection PATH` | Read a selection JSON produced by `filter_cases.py`. |
+| `--output-dir PATH` | Set the whitespace-free archive destination. |
+| `--keep-shard-structure` | Preserve part, split, and activity directories. |
+| `--workers N` | Set concurrent downloads. |
+| `--retries N` | Set retries after a failed request. |
+| `--no-resume` | Restart incomplete `.part` files. |
+| `--overwrite` | Replace completed local archives. |
+| `--dry-run` | Preview transfers without downloading. |
+
+## Local directory layout
+
+All directories created by the downloader have names without spaces:
+
+```text
+PhysInOne_data/
+├── leaderboard/
+│   ├── video_generation/
+│   ├── future_prediction/
+│   ├── physical_properties_estimation/
+│   └── motion_transfer/
+├── 3d_assets/
+│   └── PhysicBenchmark/
+└── _download_logs/
+    └── YYYYMMDD_HHMMSS/
+        ├── downloads.tsv
+        ├── extractions.tsv
+        ├── failures.tsv
+        └── summary.json
+```
+
+Published object paths are URL-encoded by the maintained lists. The downloader
+maps them to the no-space local layout above, so shell commands do not require
+special quoting for task directories.
+
+## Resume and verification
+
+Downloads are written to `.part` files and renamed atomically after completion.
+Run the same command again to reuse completed files and resume incomplete ones.
+Use `--force` only when a completed local file must be replaced.
+
+ZIP archives are checked for unsafe paths, symbolic links, and CRC errors before
+extraction. A completion marker is written only after successful extraction.
+Detailed progress and failures are recorded under `_download_logs/`.
+
+For slower or rate-limited networks, reduce concurrency:
+
+```bash
+python scripts/download_data.py \
+  --task all \
+  --workers 2 \
+  --output-dir ./PhysInOne_data
+```
+
+## 3D asset setup
+
+The completed public release contains Unreal Engine 5.5.4 resources for 14,847
+validation scenes and 120,660 training scenes, together with the released Train
+and Val resource libraries. The download contains 4,585 files and 285 ZIP
+archives (approximately 139.67 GiB before extraction). The 273 validation
+origins reserved for the Motion Transfer test set are intentionally excluded to
+prevent evaluation leakage. The release also includes ten Movie Render Queue
+configurations at `Content/PhysInOne/Scripts/`, corresponding to
+`/Game/PhysInOne/Scripts/` in Unreal Engine. Windows is recommended for the
+simplest setup; Linux is also supported with additional Unreal Engine
+configuration.
+
+Every 3D asset ZIP is extracted in place into its current parent directory. The
+Scene and Trajectory archives are deliberately kept separate:
+
+| Archive pattern | ZIP member layout | Result after extraction |
+| --- | --- | --- |
+| `Content/PhysInOne/Scenes/val_scenes_batch_*.zip` | `Val/<Physics>/<Scene>.umap` | `Content/PhysInOne/Scenes/Val/<Physics>/<Scene>.umap` |
+| `Content/PhysInOne/Scenes/train_scenes_batch_*.zip` | `Train/<Physics>/<Scene>.umap` | `Content/PhysInOne/Scenes/Train/<Physics>/<Scene>.umap` |
+| `Content/PhysInOne/Trajectories/val_trajectory_batch_*.zip` | `Val/<Physics>/<Scene>_trajectory.uasset` | `Content/PhysInOne/Trajectories/Val/<Physics>/<Scene>_trajectory.uasset` |
+| `Content/PhysInOne/Trajectories/train_trajectory_batch_*.zip` | `Train/<Physics>/<Scene>_trajectory.uasset` | `Content/PhysInOne/Trajectories/Train/<Physics>/<Scene>_trajectory.uasset` |
+
+The Train Scene and Trajectory series each span batches `0000–0120`; the Val
+series each span batches `0000–0014`. The final Train batch contains 660
+scenes, and the final Val batch contains 847 scenes.
+
+The Train and Val ZIP files can therefore share the same remote directory
+without colliding: their members begin with different split directories.
+
+After the default 3D asset download finishes, launch:
+
+```text
+PhysInOne_data/assets/PhysicBenchmark/PhysInOne.uproject
+```
+
+Do not rename directories inside `PhysicBenchmark/`. Unreal Engine resolves
+assets by project-relative package paths, including `BackGrounds_raw`,
+`__ExternalActors__`, and `__ExternalObjects__`.

Utils/metadata/repo_assignment.txt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:380a0ba378cd4c2e89b1fc6b6c26c5adf01e65f2c1c1a0c62862d360c47091ab
-size 21608478
+oid sha256:266ba0fec1c14e9990ef87c243cd715dd90c5b8235ab71faebbe8a03a4f6bff1
+size 21640129

Utils/metadata/repo_map.json

@@ -14,5 +14,6 @@
   "physinone_part13": "PhysInOneP13/PhysInOneP13",
   "physinone_part14": "PhysInOneP14/PhysInOneP14",
   "physinone_part15": "PhysInOneP15/PhysInOneP15",
-  "physinone_part16": "PhysInOneP16/PhysInOneP16"
+  "physinone_part16": "PhysInOneP16/PhysInOneP16",
+  "physinone_val": "vLAR/PhysInOne"
 }

Utils/scripts/download_lists/repo_assignment.txt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:380a0ba378cd4c2e89b1fc6b6c26c5adf01e65f2c1c1a0c62862d360c47091ab
-size 21608478
+oid sha256:266ba0fec1c14e9990ef87c243cd715dd90c5b8235ab71faebbe8a03a4f6bff1
+size 21640129

Utils/scripts/download_lists/repo_map.json

@@ -14,5 +14,6 @@
   "physinone_part13": "PhysInOneP13/PhysInOneP13",
   "physinone_part14": "PhysInOneP14/PhysInOneP14",
   "physinone_part15": "PhysInOneP15/PhysInOneP15",
-  "physinone_part16": "PhysInOneP16/PhysInOneP16"
+  "physinone_part16": "PhysInOneP16/PhysInOneP16",
+  "physinone_val": "vLAR/PhysInOne"
 }

Utils/scripts/filter_cases.py

@@ -11,8 +11,8 @@ The script parses:
     phenomena: abbreviations before "__bg..."
     background: bgXXX
     hash: final six-character code
-    part_id: e.g. physinone_part1
-    hf_zip_path: expected zip path inside the shard repo, e.g.
+    part_id: repository-map key, e.g. physinone_part1 or physinone_val
+    hf_zip_path: expected zip path inside the mapped dataset repository, e.g.
         Train/DoublePhysics/AccelConcaveSpin_AccelSurfaceSpin__bg070__K5ER39_trajectory.zip
 """