Skip to content

Add TruckDrive dataset support#22

Merged
stepankonev merged 1 commit into
mainfrom
add-truckdrive-dataset
Jun 6, 2026
Merged

Add TruckDrive dataset support#22
stepankonev merged 1 commit into
mainfrom
add-truckdrive-dataset

Conversation

@stepankonev

Copy link
Copy Markdown
Owner

TruckDrive (Torc Robotics / Princeton, CVPR 2026) is a long-range highway heavy-truck dataset. Add a first-class source-dataset processor + converter covering cameras, the ego trajectory, the Aeva FMCW LiDAR, and ego-frame 3D detections.

  • src_datasets/truckdrive/: processor, converter, IO (extracted devkit layout, gt_trajectory parsing, sync-id cross-sensor matching, float64x11 Aeva xyz decode) and calibration geometry (TF-tree BFS, pose assembly).
  • enums.py: map each rig camera to the canonical CameraDirection that matches its facing (FRONT, FRONT_LEFT, SIDE_LEFT, REAR_LEFT, ...) so keys stay consistent across datasets, and add members only for the four genuinely-extra views the eight canonical slots cannot name -- the forward telephoto pair (FRONT_{LEFT,RIGHT}NARROW) and the rear-facing side pair (SIDE{LEFT,RIGHT}_BACK). Additive, non-breaking.
  • Detections are emitted in the ego (velodyne) frame with the ego vehicle's own cab/trailer and DontCare groups excluded; the class taxonomy is vendored from the paper's supplementary Table 2.
  • configs/truckdrive.yaml, scripts/extract_truckdrive.sh and prepare_dataset_truckdrive.sh, unit + env-gated real-frame tests, and a README support-table row.

Radar, accumulated-depth, short-range Ouster LiDAR and lane lines have no StandardE2E target yet and are not ingested.

TruckDrive (Torc Robotics / Princeton, CVPR 2026) is a long-range highway
heavy-truck dataset. Add a first-class source-dataset processor + converter
covering cameras, the ego trajectory, the Aeva FMCW LiDAR, and ego-frame 3D
detections.

- src_datasets/truckdrive/: processor, converter, IO (extracted devkit
  layout, gt_trajectory parsing, sync-id cross-sensor matching, float64x11
  Aeva xyz decode) and calibration geometry (TF-tree BFS, pose assembly).
- enums.py: map each rig camera to the canonical CameraDirection that matches
  its facing (FRONT, FRONT_LEFT, SIDE_LEFT, REAR_LEFT, ...) so keys stay
  consistent across datasets, and add members only for the four genuinely-extra
  views the eight canonical slots cannot name -- the forward telephoto pair
  (FRONT_{LEFT,RIGHT}_NARROW) and the rear-facing side pair
  (SIDE_{LEFT,RIGHT}_BACK). Additive, non-breaking.
- Detections are emitted in the ego (velodyne) frame with the ego vehicle's
  own cab/trailer and DontCare groups excluded; the class taxonomy is
  vendored from the paper's supplementary Table 2.
- configs/truckdrive.yaml, scripts/extract_truckdrive.sh and
  prepare_dataset_truckdrive.sh, unit + env-gated real-frame tests, and a
  README support-table row.

Radar, accumulated-depth, short-range Ouster LiDAR and lane lines have no
StandardE2E target yet and are not ingested.
@codecov

codecov Bot commented Jun 6, 2026

Copy link
Copy Markdown

@stepankonev stepankonev merged commit eb075d6 into main Jun 6, 2026
2 of 3 checks passed
@stepankonev stepankonev deleted the add-truckdrive-dataset branch June 6, 2026 09:42
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant