JocelynW/NFF

⚠️ PRIVATE / RESTRICTED — DO NOT REDISTRIBUTE

This repository is a DERIVED work of NVIDIA PhysicalAI-Autonomous-Vehicles (NCore), which is governed by the NVIDIA Autonomous Vehicle Dataset License (nvidia-av-dataset). That licence prohibits hosting or distributing the dataset or derivative works, in whole or in part. This repo is therefore kept PRIVATE as internal research storage only. It must not be made public, shared, or redistributed without written permission from NVIDIA (see the upstream dataset card → contact). Anyone granted access must already hold their own accepted NVIDIA AV Dataset Licence.

To obtain the data legitimately, third parties should use the open nff-toolkit to download the listed sequences directly from NVIDIA and rebuild this format locally — no data is redistributed.

---

NFF — NVIDIA Foggy Field

The first multimodal foggy driving dataset combining real atmospheric fog with 7 cameras + 128-beam LiDAR (per-point µs timestamps) + 8-stream imaging radar, all with offline-refined ego-poses validated for NeRF / 3DGS / inverse rendering.

NFF-15 (Belgium, foggy highway) — 7-camera surround + LiDAR BEV with 3D boxes.

Same clip — radar BEV: returns coloured by radial (Doppler) velocity, with LiDAR (grey) and 3D boxes (green).

The 15 scenarios (front-wide preview)

Clips	15 (NFF-01…15), ~19.9 s each, 10 Hz
Total	~5.0 min · ~5.1 km driving · 14 day + 1 night
Countries	USA (4), Sweden (8), Belgium (2), Portugal (1)
Cameras	7 × ftheta fisheye → rectified pinhole (raw MP4 + pinhole JPG both kept)
LiDAR	128-beam, 10 Hz, [N,6] = x,y,z,intensity,per-point-ts(µs),ring
Radar	8 streams (4 corner SRR + front LRR + front MRR + 2 rear MRR)
Labels	3D boxes (ego frame) + 2D per-camera projections + tracks
Poses	offline-refined INS, 10 Hz, 4×4 T_rig_world

Per-clip layout

NFF-XX__<uuid8>/
  images/<7 cameras>/frame_0000.jpg ... frame_0198.jpg   # rectified pinhole
  videos/<7 cameras>.mp4 (+ .timestamps.parquet)         # raw fisheye
  lidar/sweep_0000.npy ... sweep_0198.npy                 # [N,6] float32
  radar/<stream>/...                                      # 8 streams
  labels/obstacles_3d_ego.parquet, obstacles_2d_<cam>.parquet ×7, class_taxonomy.json
  masks/<7 cameras>/                                      # valid-pixel masks
  calib/  K_rectified.json, T_camera_rig.json, T_lidar_rig.json, T_rig_world.parquet,
          camera_intrinsics_ftheta.json, lidar_intrinsics.json, lidar_timestamps.parquet
  preview/ bev_lidar_with_boxes.mp4, radar_bev_viz.mp4
  provenance.json

See manifest.csv / DATASET_SUMMARY.md for the per-clip table and _evaluation/ for full statistics and the cross-dataset comparison.

Tooling

Download / convert / visualize / load code: nff-toolkit (Apache-2.0, code only).

Citation

TBD pending paper acceptance. Provisional:
Wang, Y. et al. "NFF: NVIDIA Foggy Field — a multimodal foggy driving dataset
for 3D reconstruction under fog." WMG, University of Warwick, 2026.

Derived from NVIDIA PhysicalAI-Autonomous-Vehicles (NCore). Please also cite NCore.