---
pretty_name: MMPD-Bench
license: cc-by-nc-4.0
tags:
- polarimetry
- mueller-matrix
- biomedical-imaging
- healthy-bone-cell
- scientific-computing
- parquet
task_categories:
- image-to-image
configs:
- config_name: healthy_bone_cell
  default: true
  data_files:
  - split: train
    path: data/train-*.parquet
  - split: validation
    path: data/validation-*.parquet
  - split: test
    path: data/test-*.parquet
- config_name: external_waveplate
  data_files:
  - split: external_waveplate
    path: data/external_waveplate-*.parquet
- config_name: external_spectral
  data_files:
  - split: external_spectral_610
    path: data/external_spectral_610-*.parquet
  - split: external_spectral_650
    path: data/external_spectral_650-*.parquet
  - split: external_spectral_690
    path: data/external_spectral_690-*.parquet
---

# MMPD-Bench

## Dataset Summary

MMPD-Bench is a polarimetric imaging benchmark for learning mappings from
Mueller matrix observations to polarimetric decomposition modalities. Each
sample contains a channel-first Mueller matrix tensor and a channel-first target
tensor with six Lu-Chipman reference modalities.

Current Hugging Face release status:

- Uploaded: external waveplate test data at 633 nm.
- Uploaded: external spectral test data at 610, 650, and 690 nm.
- Uploaded: healthy bone cell `train`, `validation`, and `test` splits.

Because the waveplate tensors are 200 x 200 while the healthy bone cell and
spectral tensors are 256 x 256, the data is published as separate configs:

- `healthy_bone_cell`
- `external_waveplate`
- `external_spectral`

## Task Definition

The task is modality fission from a Mueller matrix tensor to six polarimetric
target modalities. It is not a segmentation or classification dataset.

- Input: Mueller matrix tensor, shape `[16, H, W]`, channel-first.
- Output: target modality tensor, shape `[6, H, W]`, channel-first.
- Target channel order: `D`, `Delta`, `eta`, `theta`, `psi`, `R`.

## Data Sources

This release contains healthy bone cell data from `polarization_v2` and external
test data from `polarization_v3`:

- Healthy bone cell data: source-provided patch splits from 53 sample folders.
- Waveplate data: `hwp633` and `qwp633`, measured at 633 nm.
- Multi-wavelength spectral data: selected wavelengths from `mwl_530_690`,
  currently 610, 650, and 690 nm.

## File Structure

```text
MMPD-Bench/
├── README.md
├── data/
│   ├── external_waveplate-00000-of-00001.parquet
│   ├── external_spectral_610-00000-of-00001.parquet
│   ├── external_spectral_650-00000-of-00001.parquet
│   ├── external_spectral_690-00000-of-00001.parquet
│   ├── train-00000-of-00094.parquet
│   ├── validation-00000-of-00012.parquet
│   └── test-00000-of-00011.parquet
└── metadata/
    ├── acquisition_info.json
    ├── channel_order.json
    ├── healthy_bone_cell_manifest.jsonl
    ├── healthy_bone_cell_manifest_summary.json
    ├── parameter_ranges.json
    ├── schema.json
    └── split_summary.json
```

## Tensor Schema

Common columns:

```python
{
    "sample_id": str,
    "source_id": str,
    "split": str,
    "subset": str,          # healthy_bone_cell, waveplate, or spectral
    "specimen_type": str,   # healthy_bone_cell, waveplate, or spectral
    "wavelength_nm": int | None,
    "source_path": str,
    "mueller_shape": list[int],
    "target_shape": list[int],
    "mueller": array,       # float32, channel-first
    "target": array,        # float32, channel-first
}
```

Waveplate-specific columns:

```python
{
    "plate_type": str,      # hwp or qwp
    "angle_label": str,     # e.g. 0deg, n22, p45
    "angle_deg": float,
}
```

Patch-based columns for healthy bone cell and spectral rows:

```python
{
    "patch_id": str,
    "target_encoding": str, # png_uint8_normalized_to_float32_0_1
}
```

Current tensor shapes:

- `healthy_bone_cell`: `mueller = [16, 256, 256]`,
  `target = [6, 256, 256]`.
- `external_waveplate`: `mueller = [16, 200, 200]`, `target = [6, 200, 200]`.
- `external_spectral_*`: `mueller = [16, 256, 256]`,
  `target = [6, 256, 256]`.

## Channel Conventions

Mueller channel order:

```text
M11, M12, M13, M14,
M21, M22, M23, M24,
M31, M32, M33, M34,
M41, M42, M43, M44
```

Target channel order:

```text
D, Delta, eta, theta, psi, R
```

Local source files may use names such as `Ita`, `ita`, or `Eta`; the public
channel name is normalized to `eta`.

## Physical Parameter Definitions

Mueller matrix elements are generally expected to lie within `[-1, 1]` after
normalization. In measured data, small deviations outside this range may occur
because of acquisition noise, calibration differences, numerical processing, or
normalization error. Users should inspect the value distribution for their split
and apply task-appropriate preprocessing before training, such as clipping,
standardization, or normalization based on the training set.

The target tensor follows this channel order and nominal parameter range:

```text
D, Delta: [0, 1]
eta, R: [0, pi)
theta, psi: [-pi/2, pi/2)
```

Important encoding note:

- Waveplate target arrays are stored from the source `.npy` files as float32.
- Healthy bone cell and spectral target arrays were converted from grayscale PNG
  files to float32 values normalized to `[0, 1]`; see `target_encoding`.
- Mueller matrix tensors are stored as measured/processed values, not forcibly
  clipped to `[-1, 1]`.

### Optional Mapping From Grayscale Targets to Physical Ranges

For rows whose `target_encoding` is
`png_uint8_normalized_to_float32_0_1`, the stored target tensor is a normalized
grayscale representation in `[0, 1]`. To map these values back to the nominal
physical parameter ranges used in the paper, apply:

```python
import numpy as np

TARGET_CHANNELS = ["D", "Delta", "eta", "theta", "psi", "R"]


def normalized_modalities_to_physical(target, channel_axis=0, clip=False):
    """Map normalized grayscale modalities to nominal physical ranges.

    Use this only for targets encoded as
    ``png_uint8_normalized_to_float32_0_1``. If a split already stores physical
    Lu-Chipman values, do not apply this conversion again.
    """
    target = np.asarray(target, dtype=np.float32)
    values = np.moveaxis(target, channel_axis, 0)
    if values.shape[0] != 6:
        raise ValueError(f"Expected 6 target channels, got shape {target.shape}")

    g = np.clip(values, 0.0, 1.0) if clip else values
    physical = np.empty_like(g, dtype=np.float32)
    physical[0] = g[0]                      # D: [0, 1]
    physical[1] = g[1]                      # Delta: [0, 1]
    physical[2] = np.pi * g[2]              # eta: [0, pi)
    physical[3] = np.pi * (g[3] - 0.5)      # theta: [-pi/2, pi/2)
    physical[4] = np.pi * (g[4] - 0.5)      # psi: [-pi/2, pi/2)
    physical[5] = np.pi * g[5]              # R: [0, pi)
    return np.moveaxis(physical, 0, channel_axis)
```

The inverse mapping is:

```text
D_gray = D
Delta_gray = Delta
eta_gray = eta / pi
theta_gray = theta / pi + 0.5
psi_gray = psi / pi + 0.5
R_gray = R / pi
```

Visualization note: after applying this optional physical-range mapping, use the
nominal physical ranges for color scales when comparing samples or models:
`D/Delta` in `[0, 1]`, `eta/R` in `[0, pi]`, and `theta/psi` in
`[-pi/2, pi/2]`. Per-sample min/max color scales are useful for inspection, but
they can make cross-sample or cross-modality comparisons visually misleading.
The helper script `scripts/test2.py` demonstrates both normalized targets and
physical targets with fixed physical colorbar ranges.

## Reference Label Generation

The target modalities are generated using Lu-Chipman decomposition from measured
Mueller matrices. They should be interpreted as physics-solver reference labels
for benchmarking surrogate models and physics consistency, not as direct human
annotations or absolute biological ground truth.

## Splits

| Split | Config | Subset | Samples | Shape | Notes |
|---|---|---:|---:|---|---|
| train | healthy_bone_cell | healthy_bone_cell | 6006 | `[16, 256, 256] -> [6, 256, 256]` | 94 shards |
| validation | healthy_bone_cell | healthy_bone_cell | 713 | `[16, 256, 256] -> [6, 256, 256]` | 12 shards |
| test | healthy_bone_cell | healthy_bone_cell | 643 | `[16, 256, 256] -> [6, 256, 256]` | 11 shards |
| external_waveplate | external_waveplate | waveplate | 18 | `[16, 200, 200] -> [6, 200, 200]` | 633 nm HWP/QWP |
| external_spectral_610 | external_spectral | spectral | 165 | `[16, 256, 256] -> [6, 256, 256]` | 610 nm |
| external_spectral_650 | external_spectral | spectral | 165 | `[16, 256, 256] -> [6, 256, 256]` | 650 nm |
| external_spectral_690 | external_spectral | spectral | 165 | `[16, 256, 256] -> [6, 256, 256]` | 690 nm |

## Benchmark Protocols

Evaluation configs:

1. Healthy bone cell benchmark: use config `healthy_bone_cell`, splits `train`,
   `validation`, and `test`.
2. External waveplate evaluation: use config `external_waveplate`, split
   `external_waveplate`.
3. External spectral evaluation: use config `external_spectral`, then evaluate
   `external_spectral_610`, `external_spectral_650`, and
   `external_spectral_690`.

## Loading Instructions

Install the Hugging Face datasets package:

```bash
pip install datasets
```

Load one external spectral split:

```python
from datasets import load_dataset
import numpy as np

ds = load_dataset(
    "parquet",
    data_files={
        "external_spectral_610": (
            "hf://datasets/HY2333/MMPD_Bench/"
            "data/external_spectral_610-*.parquet"
        )
    },
    split="external_spectral_610",
)

row = ds[0]
mueller = np.array(row["mueller"], dtype=np.float32)
target = np.array(row["target"], dtype=np.float32)

print(row["sample_id"])
print(mueller.shape)
print(target.shape)
```

Load via dataset config:

```python
from datasets import load_dataset

healthy = load_dataset("HY2333/MMPD_Bench", "healthy_bone_cell")
spectral = load_dataset("HY2333/MMPD_Bench", "external_spectral")
waveplate = load_dataset("HY2333/MMPD_Bench", "external_waveplate")
```

Note: in some environments, streaming reads of large nested Parquet tensors can
trigger a PyArrow shutdown issue after successful iteration. For a stable smoke
test, use non-streaming loading on a single split as shown above.

## Ethics and Limitations

The current public release focuses on healthy bone cell and external
physical/spectral evaluation data. Diseased biological samples are not included
in this release.

The targets are Lu-Chipman reference outputs. Evaluation should be interpreted
as agreement with a physics-solver reference and related physics consistency,
not as proof of absolute biological ground truth.

Measured Mueller matrix entries may be slightly outside the nominal `[-1, 1]`
range. This is expected for real acquisition pipelines; users should decide
whether to clip, standardize, or otherwise normalize values according to their
training protocol.

## License

This dataset is released under CC BY-NC 4.0.

## Citation

TODO: Add the MMPD-Bench paper citation and BibTeX entry.

## Contact

TODO: Add maintainer contact details.