Published July 30, 2026 | Version v1.0.0

HySoLiD-Drive: A Labeled Ground and Road Segmentation Benchmark for Hybrid Solid-State LiDAR (Seyond Falcon Kinetic)

  • 1. ROR icon Universidad Carlos III de Madrid

Contributors

Description

HySoLiD-Drive is the first public, labeled ground- and road-segmentation benchmark captured with an automotive hybrid solid-state (HSS) LiDAR ( the Seyond / Innovusion Falcon Kinetic)  mounted on two different vehicles. HSS sensors scan very differently from spinning ring LiDAR (non-repetitive, forward-facing, with concave/convex scan lines), and until now no public HSS benchmark for these tasks existed, so methods could only be evaluated on borrowed ring-LiDAR data. HySoLiD-Drive fills that gap.

Contents:
- 100 labeled evaluation frames, balanced 50/50 across two vehicles (different sensor height and tilt), ~73,000--95,000 points per frame (mean 88000; 8.8 million labeled points total). Distributed as full-field ASCII PLY preserving all original sensor fields (x, y, z, timestamp, intensity, flags, elongation, scan_id, scan_idx, second-return) plus a per-point label.
- 7,000 unlabeled continuous frames from two real-traffic drives (one peri-urban, one open-road + urban; 3,500 each), distributed as compressed NPZ, for geometric, self-/semi-supervised and domain-adaptation research and to reproduce the timing benchmark.

Labels (4 classes): 0 = non-ground, 1 = ground (non-road), 2 = other road/carriageway (visible but not ego-reachable), 3 = ego road.

Two evaluation tasks are derived: ground segmentation (label >= 1) and ego-road segmentation (label == 3, with class 2 excluded from the metric as an ignore region).

All point clouds are provided in an ego-centered, gravity-aligned frame (per-vehicle pitch and roll levelled). No imagery, no GPS or global coordinates, no personal data.

Documentation, evaluation code (per-point IoU / Precision / Recall / F1 scorer and data loader), reference baseline results and the leaderboard are on GitHub: https://github.com/martinpalor/HySoLiD-Drive

License: data CC BY-NC-SA 4.0; evaluation code MIT.

The dataset accompanies the method GHySSeg (Real-Time Geometric Ground and Road Segmentation for Hybrid Solid-State LiDAR).

Files

hyss_labeled_pointclouds.zip

Files (11.2 GB)

Name Size
md5:e5306f25bfdc53c6ea16f41608ce132d
284.8 kB Preview Download
md5:d3d9cde8646b38c506ad6e3dfe38451f
282.1 MB Preview Download
md5:fc211b1eb8d8f2576873f062d7751abd
10.9 GB Preview Download

Additional details

Software

Repository URL
https://github.com/martinpalor/HySoLiD-Drive
Programming language
Python , Markdown
Development Status
Active