yutiangu/HUBT_from_a_drones_perspective

OpenHUTB-CARLA UAV Small-Object Multimodal Dataset

This is a synthetic UAV-view dataset collected in OpenHUTB/CARLA Town10HD. It contains 500 synchronized frames for small-object detection and multimodal perception experiments.

Dataset summary

Item	Value
Frames	500
Annotations	1,112
Resolution	1920 x 1080
Classes	vehicle, pedestrian
Map	Carla/Maps/Town10HD
Camera FOV	55 degrees
Train/validation/test	350 / 100 / 50

The split frame IDs are disjoint. Each geometry frame is rendered with one weather selected from a balanced random schedule.

Modalities

• RGB visible-light images
• Metric camera depth and projected LiDAR depth
• Camera-space surface normals
• Semantic segmentation IDs and color previews

The RGB, depth, surface-normal, segmentation, LiDAR, and annotation files for a frame are spatially aligned and synchronized.

Weather conditions

• ClearNoon
• ClearSunset
• ClearNight
• FoggyNoon
• SnowNoon
• DustStorm

FoggyNoon includes depth-dependent visibility attenuation. SnowNoon includes deterministic falling-snow rendering but does not simulate accumulated snow on surfaces.

Annotations

The dataset provides:

• YOLO detection labels in paired_weather/seq_0000/labels_yolo/
• Per-frame JSON annotations in paired_weather/seq_0000/annotations/
• COCO annotations in coco/
• Reproducible image lists in splits/

YOLO class IDs are zero-based (0=vehicle, 1=pedestrian). COCO category IDs are one-based. Bounding boxes are derived from OpenHUTB/CARLA actor coordinates and filtered using semantic and metric-depth visibility. Actors with less than 50 percent visible area are not annotated.

Directory overview

paired_weather/seq_0000/
  rgb/<weather>/
  depth/
    npy/
    vis_16bit/
    color/
    lidar/
  surface_normal/
    png/
    npy/
  segmentation/
    id/
    color/
  annotations/
  labels_yolo/
coco/
splits/
splits_by_weather/

See DATASET_README.md, dataset_manifest.json, and quality_report.json for the complete format, sensor settings, and quality statistics.

Known limitations

• All frames are synthetic and use one CARLA map, so same-domain scores do not measure real-world or cross-map generalization.
• The dataset has no empty frames and contains only two detection classes.
• Severe occlusions are filtered by the annotation visibility policy.
• Surface normals and semantic segmentation are simulator-derived ground truth; they should be treated as privileged supervision unless an equivalent input is available at deployment time.

License

No standalone dataset license has been declared yet. Users must also comply with the applicable OpenHUTB, CARLA, Unreal Engine, and map/asset terms.