README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- image-classification
- depth-estimation
language:
- en
tags:
- autonomous-driving
- polarization
- polarimetric-imaging
- road-surface
- multi-modal
- lidar
- benchmark
pretty_name: PRISM
size_categories:
- 10K<n<100K
---

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

> **Anonymous submission to NeurIPS 2026 Evaluations & Datasets Track.**
> Author identities and the camera-ready release URL will be revealed at the
> camera-ready stage.

PRISM is a polarimetric road-surface dataset and benchmark of **47,098
time-synchronized frames** combining trichromatic linear polarization,
co-boresighted RGB, 128-channel LiDAR, and RTK-GNSS/INS, captured from an
in-vehicle sensor rig across proving-ground and open-road environments under
clear, overcast, rainy, foggy, and snowy conditions. It is designed as an
evaluative testbed for measuring whether, and under which conditions,
polarimetric channels add discriminative signal to RGB on road-surface
perception.

## Why PRISM

RGB cameras struggle on the road surface itself: pavement is nearly
texture-less at the centimetre scale, and dry, wet, and icy asphalt can look
photometrically identical. Polarization is the natural complement, since
Fresnel optics ties the angle of linear polarization (AoLP) to surface-normal
azimuth and the degree of linear polarization (DoLP) to refractive index,
but no public dataset has so far paired visible-light polarization with the
geometric and semantic ground truth required to test whether this physical
promise translates into measurable gains on real roads.

## Dataset at a glance

| Property | Value |
|---|---|
| Frames | 47,098 (time-synchronized, multi-modal) |
| Sessions (datasets) | 41, anonymized as `0106_dataset`, `0112_dataset`, … |
| Modalities | RGB, four-orientation linear polarization (0°/45°/90°/135°), accumulated LiDAR scan, vehicle state |
| Image resolution | 2448 × 2048, 12-bit |
| LiDAR | Ouster OS1, 128 channels |
| Pose | RTK-corrected, tightly-coupled GNSS/INS |
| Synchronization | GPS-PPS hardware trigger |
| Environments | Proving-ground (35.2%), Open-road urban/rural/highway (64.8%) |
| Weather | Clear (65.1%), Rainy (14.7%), Overcast (13.5%), Foggy (5.0%), Snow (1.7%) |
| Materials | Asphalt, Concrete, Belgian block, Gravel, Other |
| Surface states | Dry, Damp, Wet, Slush, Snow-covered |
| Elevation supervision | Derived from `lidar_accum_scan/*.pcd` (LiDAR-inertial SLAM-accumulated point clouds) |
| Total size | ≈ 1.59 TB |
| License | CC-BY-NC-SA 4.0 |

## Sensor platform

- **Color-polarization camera**: Lucid Triton 5.0 MP, Sony IMX264MYR, on-chip 2×2 polarizer super-pixels at {0°, 45°, 90°, 135°} over a Bayer CFA.
- **RGB camera**: Lucid Triton 5.0 MP, Sony IMX264 (same sensor family without polarizer array, co-boresighted with the polarization camera).
- **LiDAR**: Ouster OS1, 128 channels.
- **GNSS/INS**: NovAtel PwrPak7, RTK-corrected, tightly-coupled.

The two cameras are drawn from the same sensor family to minimise
modality-comparison confounds. Camera–camera and camera–LiDAR extrinsics
were calibrated with the MATLAB Camera Calibration and Lidar–Camera
Calibration toolboxes, and cross-validated by comparing direct vs composed
extrinsics. Polarimetric fidelity through the front windshield was
characterised with a controlled LCD-monitor protocol (see paper, Appendix F).

## Repository layout

The dataset is distributed as **per-session ZIP files**, organised at the
top level by train / validation split:

```
PRISM-Dataset/
├── README.md
├── labels.json                          # frame-level material/state labels (whole dataset)
├── calibration/                         # camera intrinsics + inter-sensor extrinsics
│   ├── 0106_calibration.yaml
│   ├── …
│   └── 0327_calibration.yaml
├── train/
│   ├── 0106_dataset.zip
│   ├── 0112_dataset.zip
│   ├── 0119_1_dataset.zip
│   ├── …
│   └── 0327_11_dataset.zip
└── val/
    ├── 0112_dataset.zip          # last 30 % of one snow session (intra-session split)
    ├── 0119_1_dataset.zip
    └── …
```

`labels.json` and `calibration/` are kept outside the per-session ZIPs so
they can be inspected without unpacking any archive.

Some session names appear in **both** `train/` and `val/`: those are sessions
that were split *within* a recording (intra-session split) to preserve
training-set coverage of rare classes such as `slush` and `snow-covered`,
which originate from a small number of capture sites. See paper Appendix E
for the full split protocol.

Inside each ZIP, the layout is:

```
{session_anon}/                              # e.g. 0106_dataset/
├── sequence_001/
│   ├── 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
├── sequence_002/
│   └── …
└── vehicle_state/                           # session-level (NOT per-sequence)
```

**File naming.** All files share the same nanosecond Unix timestamp as the
filename stem (e.g. `1736200123456789012.png`). RGB, the four polarisation
images, and the corresponding accumulated LiDAR scan for the same instant
share a stem, which is also how the train / val split is internally applied
to intra-session sessions.

**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. The Stokes-to-AoLP/DoLP formulas
are given in the paper (Section 3) and a reference implementation is
included in the code release.

**LiDAR.** The `lidar_accum_scan/` directory contains **LiDAR-inertial
SLAM-accumulated point clouds** (one PCD per frame) produced by the
pipeline described in paper Section 4.3. These accumulated clouds are
already deskewed, aligned to a common ground frame, and ICP-refined on
static ground segments, so they can be used directly. They serve as the
**source of dense road-surface elevation supervision** for Track 2:
elevation BEV is rasterised from the PCD onto the 10 cm × 10 cm grid at
training and evaluation time. Single-sweep raw LiDAR is not released at
this time.

**Vehicle state.** `vehicle_state/` is a session-level directory (not
per-sequence) containing RTK-INS pose and synchronised vehicle-bus signals.

**Labels.** Frame-level surface material and state labels are released as
a single `labels.json` at the repository root, indexed by
`(session_id, sequence_id, timestamp)`.

**Calibration.** Camera intrinsics and inter-sensor extrinsics
(RGB ↔ polarisation, camera ↔ LiDAR) are released as YAML files under the
`calibration/` directory at the repository root.

**Privacy.** Faces and licence plates throughout the released imagery are
replaced in-place by Gaussian blur (OpenCV `cv2.GaussianBlur`, 31×31 kernel)
on the raw polariser-resolved intensities. The road region is unaffected.
See paper Appendix K.

## Anonymised session naming

To support double-blind review, the original session folder names (which
encode location and date) are remapped to chronological-anonymised names.
The full mapping is provided as `dataset_name_remapping.json` alongside the
code release, and will be revealed at the camera-ready stage. Anonymised
names follow the pattern `{date_index}_{seq_index}_dataset` where
`date_index` is a chronological index of the recording day.

## Benchmark tracks

PRISM defines two benchmark tracks, both evaluated under a shared
five-variant input ablation crossing RGB, monochromatic / trichromatic
polarisation, and their combinations:

| ID | Variant | Channels | Role |
|---|---|---|---|
| (a) | RGB | 3 | RGB-intensity baseline |
| (b) | Mono-polar | 3 | DoLP + (sin 2φ, cos 2φ) AoLP, luminance-averaged |
| (c) | Tri-polar | 9 | Per-channel R/G/B DoLP + AoLP |
| (d) | RGB + Mono-polar | 3 \| 3 | Two-stream fusion |
| (e) | RGB + Tri-polar | 3 \| 9 | Two-stream fusion (full polarimetric) |

**Track 1 — Road-surface condition classification.** Joint per-frame
prediction of surface material ∈ {asphalt, concrete, Belgian block, gravel}
and surface state ∈ {dry, damp, wet, slush, snow-covered}. Reported metrics:
accuracy, macro-precision, macro-recall, macro-F1 (primary).

**Track 2 — Road-surface elevation estimation.** Dense BEV elevation
prediction on a 10 cm × 10 cm grid covering longitudinal x ∈ [7, 12] m and
lateral y ∈ [−2, 2] m (one-lane near-field preview window for active
suspension). Reported metrics: MAE, RMSE, and the fraction of cells whose
absolute error exceeds 0.5 / 1.0 / 2.0 cm.

Both tracks are further analysed under stratification by surface state,
surface material, and weather to localise where polarimetric channels
matter most.

## Splits

| Partition | Frames |
|---|---|
| Train | ≈ 84.7 % |
| Validation | ≈ 15.3 % |
| Test | Released at camera-ready, scene-disjoint where possible |

Frames captured during surface transitions, where two or more material or
state classes are simultaneously present in a single image, are excluded
(≈ 13 % of all collected frames) to ensure unambiguous single-label
annotation. Train / validation partitions are defined at the session level
where possible, with two exceptions to preserve coverage of rare classes:
sparse special-surface sessions (gravel, dust-room, plastic-bump-only) use
within-session splitting, and single-source surface states (snow-covered,
slush) allocate the last 30 % of the session to validation. See paper
Appendix E.

## Quick start

```python
from huggingface_hub import snapshot_download
import zipfile, os, json

# 1. Pull labels, calibration, and a single train session
local = snapshot_download(
    repo_id="NeurIPS-2026-PRISM/PRISM-Dataset",
    repo_type="dataset",
    allow_patterns=[
        "labels.json",
        "calibration/*.yaml",
        "train/0106_dataset.zip",
    ],
)

# 2. Unpack the session
with zipfile.ZipFile(os.path.join(local, "train/0106_dataset.zip")) as zf:
    zf.extractall("prism_local")

# 3. Walk a sequence
seq = "prism_local/0106_dataset/sequence_001"
print(sorted(os.listdir(seq)))
# → ['lidar_accum_scan', 'polar', 'rgb']

# 4. Lookup labels for one frame
labels = json.load(open(os.path.join(local, "labels.json")))
ts = sorted(os.listdir(f"{seq}/rgb"))[0].replace(".png", "")
print(labels["0106_dataset"]["sequence_001"][ts])
# → {"material": "asphalt", "state": "dry"}
```

**Deriving elevation BEV from `lidar_accum_scan/`.**  Each `.pcd` is an
already-accumulated, ego-aligned ground-frame point cloud. To produce the
BEV elevation grid used by Track 2, project the points onto a 10 cm × 10 cm
grid covering longitudinal x ∈ [7, 12] m and lateral y ∈ [−2, 2] m and take
a per-cell statistic (mean / median) over the points falling inside each
cell, with a per-cell valid mask flagging cells with too few returns. A
reference implementation is included in the code release.

To pull the full dataset (≈ 1.59 TB), drop the `allow_patterns` argument
and ensure you have sufficient disk space.

## Intended use and out-of-scope use

**Intended.** Research on road-surface perception, polarimetric sensing,
multi-modal fusion, physics-informed learning, and shape-from-polarization
in unconstrained outdoor settings.

**Out of scope.**
- Direct deployment in safety-critical systems without further validation.
- Re-identification of individuals or vehicles. Faces and licence plates
  are detected and Gaussian-blurred (31×31 kernel) prior to release; the
  road region is unaffected.
- Night-time perception. PRISM sessions are predominantly daytime, since
  a polariser halves incoming flux and polarimetric SNR degrades at low
  illumination.
- Absolute polarimetric measurement beyond the fidelity bounds documented
  in Appendix F of the paper. Applications requiring absolute AoLP
  recovery should add per-session optical-path calibration.

## Known limitations

- **Long tail.** Rare classes (snow weather 1.7 %, snow-covered state 3.1 %,
  gravel material 2.1 %, foggy weather 5.0 %) have low absolute frequency.
  Stratified results on these classes should be read as indicative.
- **Geographic scope.** All sessions were collected in a single country;
  asphalt composition, Belgian-block prevalence, and weather distribution
  are location-specific.
- **Label granularity.** Surface material and state labels are at frame
  level, not pixel level. Pixel-level labels on a subset are a planned
  extension.
- **Ground-truth precision.** Elevation derived from accumulated LiDAR is
  strong but imperfect on specular wet surfaces and under GNSS multipath;
  absolute claims below ≈ 0.5 cm should be interpreted in light of the
  proving-ground qualitative validation in Appendix I of the paper.

## Citation

```bibtex
@inproceedings{prism2026,
  title     = {PRISM: Polarimetric Road-surface Intelligent Sensing and Measurement Dataset},
  author    = {Anonymous},
  booktitle = {Advances in Neural Information Processing Systems 39 (NeurIPS 2026), Evaluations and Datasets Track},
  year      = {2026},
  note      = {Anonymous submission}
}
```

## License and terms of use

PRISM is released under **Creative Commons Attribution-NonCommercial-ShareAlike 4.0
International (CC-BY-NC-SA 4.0)**.

By downloading or using PRISM, you agree to:

1. **Attribution.** Cite the PRISM paper in any publication that uses the dataset.
2. **Non-commercial use only.** Academic and non-commercial research use,
   derivative works, and redistribution with attribution are permitted under
   the same licence. Commercial use is not permitted without a separate
   agreement with the authors.
3. **Share-alike.** Derivative works must be released under CC-BY-NC-SA 4.0.
4. **No re-identification.** Do not attempt re-identification of any
   individual or vehicle that may inadvertently appear in the imagery, even
   though faces and licence plates have been blurred.

## Contact

For dataset-related questions during the review period, please use the
OpenReview submission discussion. Author contact will be revealed at the
camera-ready stage.