NeurIPS-2026-PRISM/PRISM-Dataset-Sample

PRISM Sample: Polarimetric Road-surface Intelligent Sensing and Measurement Dataset

Anonymous submission to NeurIPS 2026 Evaluations & Datasets Track.

This is a representative sample of the PRISM dataset, designed to enable reviewers and researchers to inspect data quality without downloading the full ~1.6 TB dataset.

Why a sample dataset?

The full PRISM dataset contains 47,098 time-synchronized frames across 41 sessions. This sample provides:

• Quick quality inspection: Download < 4 GB instead of 1.6 TB
• Representative coverage: All surface types and conditions
• Reproducible sampling: Exact methodology documented in SAMPLING_MANIFEST.json

Sample at a glance

Property	Value
Frames	~30 (3 frames × 10 datasets)
Datasets	10 representative sessions
Modalities	RGB, four-orientation polarization (0°/45°/90°/135°), accumulated LiDAR, vehicle state
Image resolution	2448 × 2048, 12-bit
Surface types	Asphalt, Concrete, Belgian block, Gravel
Surface conditions	Dry, Damp, Wet, Slush, Snow-covered
Size	< 4 GB
Format	ZIP files (train.zip, val.zip)

Sampling strategy

Selection criteria

We selected 10 representative datasets from 41 total sessions to cover:

• All surface types: asphalt, concrete, belgian_block, gravel
• All road conditions: dry, damp, wet, slush, snow_covered
• Both train and validation splits: Including intra-session splits

Each dataset includes 1 sequence with 3 uniformly sampled frames, providing temporal coverage while maintaining manageable file size.

Representative datasets

Dataset	Surface Type	Condition	Split	Frames
0106	asphalt	dry	train	3
0112	asphalt	snow_covered	train (intra-split)	3
0129_1	asphalt	damp	train	3
0318_9	asphalt	wet	train	3
0124	asphalt	slush	train	3
0128_1	concrete	snow_covered	val	3
0128_3	concrete	damp	val	3
0327_3	belgian_block	dry	val	3
0328_5	belgian_block	snow_covered	val	3
0327_9	gravel	dry	val	3

Temporal sampling

For each sequence:

• 3 frames uniformly sampled across the full sequence duration
• Vehicle state data: ±100ms window around each sampled frame (~60 files per sequence @ 100Hz)
• All sensor modalities: RGB, polarimetric (0°, 45°, 90°, 135°), LiDAR accumulated scan

Privacy protection

Privacy measures identical to the full dataset:

• RGB images: Faces and licence plates replaced by Gaussian blur (OpenCV cv2.GaussianBlur, 31×31 kernel)
• Polarimetric images: Masked where corresponding RGB masks exist
• Vehicle state: GPS coordinates included (public roads only)

Repository layout

The sample dataset is distributed as two ZIP files matching the full dataset structure:

PRISM-Dataset-Sample/
├── README.md                            # This file
├── train.zip                            # Training split samples
└── val.zip                              # Validation split samples

Inside each ZIP (train.zip or val.zip):

train/  (or val/)
├── 0106/                                # Anonymized session name
│   ├── sequence_006/
│   │   ├── rgb/                         # *.png   RGB images
│   │   ├── polar/
│   │   │   ├── 0d/                      # *.png   polariser at  0°
│   │   │   ├── 45d/                     # *.png   polariser at 45°
│   │   │   ├── 90d/                     # *.png   polariser at 90°
│   │   │   └── 135d/                    # *.png   polariser at 135°
│   │   └── lidar_accum_scan/            # *.pcd   accumulated scan
│   └── vehicle_state/                   # *.txt   session-level (NOT per-sequence)
├── 0124/
└── ...

File naming. All files share the same timestamp as filename stem (e.g., 1736200123_456.png for images, 1736200123_456.pcd for LiDAR, 1736200123_456.txt for vehicle state). The format is {seconds}_{milliseconds} derived from Unix nanosecond timestamps.

Polarisation. PRISM ships raw four-orientation polariser-resolved intensities rather than pre-computed Stokes / AoLP / DoLP maps. This keeps the release closer to the sensor and lets users compute polarimetric quantities under their own conventions.

LiDAR. The lidar_accum_scan/ directory contains LiDAR-inertial SLAM-accumulated point clouds (one PCD per frame). These accumulated clouds are already deskewed, aligned to a common ground frame, and ICP-refined on static ground segments.

Vehicle state. vehicle_state/ is a session-level directory (not per-sequence) containing RTK-INS pose and synchronised vehicle-bus signals at 100 Hz. Each .txt file contains 29 comma-separated values:

timestamp, latitude, longitude, altitude,
roll, pitch, yaw,
velocity_x, velocity_y, velocity_z,
acceleration_x, acceleration_y, acceleration_z,
angular_velocity_x, angular_velocity_y, angular_velocity_z,
... (additional vehicle dynamics data)

Full dataset

This sample represents approximately 0.06% of the full PRISM dataset.

Property	Full Dataset	Sample
Sessions	41	10
Frames	47,098	~30
Size	~1.6 TB	<4 GB
Coverage	All conditions	Representative conditions

Full dataset: https://huggingface.co/datasets/NeurIPS-2026-PRISM/PRISM-Dataset

The script applies uniform temporal sampling to each selected sequence and copies files according to the per-file masked priority logic documented in the full dataset README.

License

CC-BY-NC-SA 4.0

Citation

If you use this dataset, please cite:

@article{prism2026,
  title={{PRISM}: Polarimetric Road-surface Intelligent Sensing and Measurement Dataset},
  author={Anonymous},
  journal={NeurIPS Datasets and Benchmarks Track},
  year={2026}
}

Contact

For questions about this sample dataset or the full PRISM dataset, please open an issue on the dataset repository during the review period.