LEAD-DVS

LEAD expert-driving data collected in CARLA 0.9.15 with synchronized RGB, depth, semantic/instance segmentation, LiDAR, radar, HD map, metadata, 3D bounding boxes, and a forward-facing DVS event camera.

Code release

Dataset construction, preprocessing, training, and evaluation code will be published in SamanthaZhang-stu/ReflexWorldModel. This GitHub repository is the designated code-release location for the project.

Contents

• 1,579 successful expert routes across 43 scenario categories.
• 892 routes have status Completed; 687 have status Perfect.
• 4,205,713 files before archiving, approximately 188 GiB on disk.
• DVS resolution: 768 x 768.
• DVS thresholds: positive 0.3, negative 0.3.
• DVS events are capped at 40,000 per saved window.

The dataset is distributed as one uncompressed tar archive per scenario under archives/. Each archive restores paths rooted at data/<scenario>/.... Uncompressed tar is intentional: most payloads are already compressed, and it avoids expensive recompression while retaining the original directory layout.

Restore

Download all files under archives/, then run:

mkdir -p LEAD-DVS
for archive in archives/*.tar; do
  tar -xf "$archive" -C LEAD-DVS
done

The restored layout is:

data/<scenario>/<route>/
  bboxes/
  depth/
  depth_perturbated/
  depth_raw/
  dvs/
  hdmap/
  hdmap_perturbated/
  instance/
  instance_perturbated/
  lidar/
  metas/
  radar/
  radar_perturbated/
  rgb/
  rgb_perturbated/
  semantics/
  semantics_perturbated/
  results.json

See entries.tsv for the route index and archives.sha256 for archive integrity checks.

Splits and dataset viewer

The Hub viewer reads the route-level JSONL indexes under splits/; it does not interpret the raw scenario tar files as WebDataset shards. Each viewer row describes one complete route and points to its source archive and member path.

The split follows LEAD's Town13-held-out evaluation principle:

• test: all 539 Town13 routes (fully held out from training).
• validation: 105 non-Town13 routes, selected deterministically within every scenario category by SHA-256 ranking (10%).
• train: the remaining 935 non-Town13 routes.

The exact deterministic method and per-town/per-scenario counts are recorded in splits/split_summary.json.