PolySight Dataset

Polarimetric bistatic SAR dataset collected with a 77 GHz FMCW radar (TI AWR2243). Two radar boards receive orthogonal polarizations (RH, RV) while a single TX board illuminates the scene. A 2-D motorized stage rasters the target across the aperture.

Directory Structure

data/
├── unsynced/          # Raw unsynchronized ADC captures (1 example collection)
│   └── <collection>/
│       ├── RH/                   # Raw DCA1000 captures, H-pol (.bin + _log.csv)
│       ├── RV/                   # Raw DCA1000 captures, V-pol (.bin + _log.csv)
│       ├── LS/                   # Raw DCA1000 captures, linear stage (.bin + _log.csv)
│       ├── _adcData/RH/          # Decoded per-frame .npy files, H-pol
│       ├── _adcData/RV/          # Decoded per-frame .npy files, V-pol
│       ├── adcData/RH/           # Synchronized per-frame .npy files, H-pol
│       ├── adcData/RV/           # Synchronized per-frame .npy files, V-pol
│       ├── collection.log
│       ├── data_collection_cfg.json
│       ├── devices_comm_cfg.json
│       └── radar_id.json
│
├── raw/               # Synchronized radar cubes (70 collections, ~72 GB, ~1 GB each)
│   └── <collection>/
│       ├── RH_synced.npy         # Synchronized ADC cube, H-pol receiver
│       ├── RV_synced.npy         # Synchronized ADC cube, V-pol receiver
│       ├── data_collection_cfg.json
│       ├── devices_comm_cfg.json
│       └── radar_id.json
│
├── labels/            # SAR images + bounding-box labels (85 collections)
│   └── <collection>/
│       ├── sar_h.npy             # Complex SAR image, H-pol (150x150, complex64)
│       ├── sar_v.npy             # Complex SAR image, V-pol (150x150, complex64)
│       └── <collection>.json     # Label file (bounding boxes + SAR config)
│
├── materials/                    # Cropped material samples (see below)
│   ├── ceramic_60.pkl
│   ├── water_55.pkl
│   └── ...
│
└── README.md

Download

The polysight dataset is hosted on Hugging face: https://huggingface.co/datasets/xinghs/polysight

Files are downloaded into the current directory, preserving the original folder structure.

Folder	Size	Contents
raw/	~72 GB	70 synchronized radar cubes (~1 GB each)
labels/	~66 MB	84 SAR image collections with labels
materials/	~1 MB	72 cropped material sample files (.pkl)

Use download.py to selectively download subfolders, install the Hugging Face Hub client:

pip install huggingface_hub

# List all available folders
python download.py --list

# Download a top-level folder
python download.py materials

# Download specific collections
python download.py labels/20250603_185728_oil_water labels/20250605_194155_water

# Download with glob patterns
python download.py "labels/20250625*"

# Download multiple folders at once
python download.py materials labels raw

# Download everything
python download.py --all

---

Pipeline

unsynced/  ──[decode + sync]──>  raw/  ──[SAR generation]──>  labels/  ──[crop]──>  materials/

1. unsynced/ - Raw ADC Captures

Raw ADC frames captured by three DCA1000 FPGA boards over Ethernet. Each radar board produces independent packet streams that are not yet time-aligned. These are provided as examples for reproducing the full decode and synchronization pipeline (polysight.collection.decode). Most users can skip this stage and start from raw/.

2. raw/ - Synchronized Radar Cubes

Time-synchronized 3-D ADC data cubes, one per polarization channel:

• RH_synced.npy - Horizontal polarization receiver
• RV_synced.npy - Vertical polarization receiver

Each file is a NumPy array of shape (num_positions, num_antennas, num_adc_samples) containing the decoded, packet-loss-corrected, and cross-board synchronized IF signal. Generated from unsynced/ data using scripts/decode.py.

3. labels/ - SAR Images + Labels

SAR images generated from raw/ cubes via GPU-accelerated FFT back-projection (scripts/generate_sar.py, notebooks/02_sar_generation.ipynb).

Each collection folder contains:

• sar_h.npy / sar_v.npy - Complex-valued SAR images (150x150, complex64)
• <collection>.json - Metadata and bounding-box labels, created with notebooks/03_labeling.ipynb

The JSON label file contains:

{
  "description": "1tx, oil, water, copper",
  "labels": {
    "oil":    [x0, y0, x1, y1],
    "water":  [x0, y0, x1, y1],
    "copper": [x0, y0, x1, y1]
  },
  "sar_config": {
    "numX": 350, "numY": 200,
    "dx": 0.002, "dy": 0.002,
    "output_size_x": 150, "output_size_y": 150,
    "chirp_configs": { "f0": 76.01e9, "slope": 38.8e12, ... },
    "antenna_configs": { ... },
    "ant_locs": [[x,y,z], ...],
    "RH_config": { "alignment_configs": { ... }, ... },
    "RV_config": { ... }
  }
}

4. materials/ - Cropped Material Samples

Pickled dictionaries containing cropped SAR pixel regions for individual materials, extracted from labeled SAR images. Each .pkl file represents one material measured at a specific incident angle. Used by the evaluation notebooks for permittivity estimation and material classification.

File naming: <material>_<angle>.pkl (e.g., ceramic_60.pkl, water_55.pkl)

Structure of each .pkl file:

{
    "name":           str,        # Material name (e.g., "ceramic")
    "thickness":      float,      # Sample thickness in meters (e.g., 0.00127)
    "collection_name": str,       # Source collection (e.g., "20250623_182841_multi")
    "incident_angle": float,      # Measurement angle in degrees (e.g., 60.0)
    "ref_name":       str,        # Reference material name (e.g., "copper")
    "ref_thickness":  float,      # Reference thickness in meters
    "values":         list[ndarray],  # 3 complex64 arrays: [H/V ratio, H, V]
                                      # each shaped (rows, cols) from bounding box
    "ref_values":     list[ndarray],  # Same structure for the reference material
}

The values and ref_values lists each contain 3 arrays corresponding to the H/V polarimetric ratio, H-pol, and V-pol SAR pixel regions cropped from the bounding box. The polarimetric ratio (element 0) is used with Fresnel equations to estimate relative permittivity via polysight.eval.permittivity.ratio_to_epsilon.

Collection Categories

Category	Count	Description
liquids	11	Oil, water, soy sauce, syrup, sugar-water series
solids	24	Plastics, wood, rubber, metals, composites
roughness	3	Wood grain orientations, sandpaper grits
ablation	9	Parameter sensitivity
concentration	7	Sugar-water concentration gradient (0-6 spoons)
ceramics	8	Alumina ceramic vs plastic at varying positions
misc	8	Oil, powder, rubber, PVC, mixed materials