# HandNet utilities

## Inspect a frame

`inspect_sample.py` is a new, standalone Python reader for the archival MAT files. It prints field shapes, dtypes, finite ranges, sensor positions and zero-based projected sensor pixels. It can export a calibrated point cloud. It does not download data or run a learned detector.

Install dependencies in your preferred Python environment:

```bash
python -m pip install numpy scipy
```

From a downloaded copy of this repository:

```bash
# Inspect the first Data*.mat member of a ZIP without extracting it.
python tools/inspect_sample.py archives/ValidationData.zip

# List available MAT members, then select an exact member or zero-based index.
python tools/inspect_sample.py archives/ValidationData.zip --list
python tools/inspect_sample.py archives/ValidationData.zip --index 12
python tools/inspect_sample.py archives/ValidationData.zip --member ValidationData/Data_0000000.mat

# Inspect an already extracted MAT file.
python tools/inspect_sample.py /path/to/Data_0000000.mat

# Export a sparse, segmented point cloud and all sensor annotations.
python tools/inspect_sample.py archives/ValidationData.zip --step 3 --pointcloud frame.npz
```

The example `--member` name is illustrative: use `--list` to obtain the archive's actual member path. `--index` selects a member in sorted filename order, not chronological order. For `TrainData.rar`, first use a RAR-capable extractor and inspect an extracted MAT file. The script deliberately does not invoke external archive executables or extract archive contents to disk.

Point-cloud output is a compressed NumPy archive with `xyz_mm` (`N,3`), `sensor_xyz_mm` (`3,6`), `sensor_rotations` (`3,3,6`), `pixel_uv` (`N,2`) and `K` (`3,3`). It retains camera axes (+X right, +Y down, +Z away). By default only nonzero depth within `segmap` is exported. Add `--include-background` to keep all nonzero depth. Point-cloud export creates the requested output file and refuses to overwrite an existing one.

The default calibration is `../calibration.json`, resolved relative to the script. Use `--calibration custom.json` for another compatible calibration file containing a `K` matrix. This reader targets the MAT format distributed by HandNet, not arbitrary MAT v7.3/HDF5 datasets.

## Reproduce the Explorer sample bundle

`extract_space_samples.py` builds the compact image/point-cloud bundle used by [HandNet Explorer](https://huggingface.co/spaces/twerdster/HandNet-Explorer). It reads MAT members directly from five source ZIPs, without extracting the full archives. Install its additional image dependency:

```bash
python -m pip install numpy scipy pillow
python tools/extract_space_samples.py --archive-dir ./archives --output-dir ./explorer-data
```

Use a new, empty output directory. The source directory must contain `TestData.zip`, `TestData_Perturbed.zip`, `ValidationData.zip`, `ValidationData_DerotGT.zip` and `TrainData_DerotGT.zip`. The original RAR is not used in this small visualization selection.

The deterministic selection contains 96 frames: 24 original test, 24 perturbed test, 16 original validation, 16 ground-truth-normalized validation, and 16 ground-truth-normalized training frames. Within each archive, it chooses evenly spaced indices in sorted member order. Corresponding test/validation variants share selected member indices; source session/frame IDs are retained for inspection.

The output contains `samples.json`, sparse little-endian float32 XYZ clouds, and PNG images for depth, IR, source labels and all six sensor heatmaps. Point clouds keep every second pixel in each axis within the nonzero-depth hand mask. Depth and IR are contrast stretched for viewing; heatmaps use a fixed 0-to-1 grayscale scale. These are display derivatives, not replacement training archives. Each sample records its source archive, member and MAT SHA-256 hash. The generated manifest uses paths relative to an Explorer root with the bundle in `data/`; copy `explorer-data/` to the app's `data/` directory when reproducing the site.

## Reproduce the Hub preview

After generating the Explorer bundle, optionally export its 96 display rows to the Hub-compatible Parquet preview:

```bash
python -m pip install datasets pyarrow pillow
python tools/build_parquet_preview.py --bundle-dir ./explorer-data --output ./preview/samples.parquet
```

The exporter embeds the depth, infrared and colorized-label PNGs, sensor poses and source identifiers; it also writes `preview/samples.schema.json`. It checks the 96-row count, Hugging Face schema metadata and an image-byte round trip. This preview does not contain every frame, the full point clouds or raw depth arrays. Use the original archives for training and quantitative image analysis.

## Licenses

The new utilities and this tool documentation use the [MIT license](LICENSE). Dataset arrays and their visualization derivatives use the repository's CC BY 4.0 license. Historical MATLAB code is linked in the dataset documentation and is not copied into these utilities.