yutiangu/HUBT_from_a_drones_perspective

HUTB From a Drone's Perspective

Dataset Summary

HUTB From a Drone's Perspective is a synthetic, multi-map UAV dataset generated with OpenHUTB/CARLA. It provides synchronized visible RGB, metric depth, surface normals, target semantic segmentation, LiDAR, and vehicle/pedestrian detection annotations from elevated oblique viewpoints.

The current release contains:

• 4,081 synchronized frames at 1920 x 1080 pixels.
• 8 simulated maps.
• 6 weather and illumination conditions.
• 16,275 trainable object instances.
• 6,707 vehicle boxes and 9,568 pedestrian boxes.
• YOLO, COCO, and per-frame JSON annotations.
• Per-map train/validation/test splits using an approximately 70/20/10 ratio.

Each frame is assigned exactly one weather condition. RGB files in different weather directories therefore represent disjoint frames rather than multiple weather renders of the same frame.

Chinese Summary

本数据集是基于 OpenHUTB/CARLA 生成的多地图无人机视角合成数据集。当前版本包含 4,081 组 1920 x 1080 同步数据，覆盖可见光 RGB、米制深度、表面法向、车辆与行人 目标语义分割、LiDAR 点云及其 RGB 对齐投影，并提供 YOLO、COCO 和逐帧 JSON 标注。 数据覆盖 8 张地图和 6 种天气，检测类别为车辆与行人。

Maps and Annotations

Map	Frames	Train	Validation	Test	Vehicles	Pedestrians
CCSP_Zhongdian_Software_Park	181	126	36	19	182	311
HutbCarlaCity	300	210	60	30	533	618
Town02_Opt	600	420	120	60	1,244	1,108
Town03_Opt	600	420	120	60	864	1,198
Town04_Opt	600	420	120	60	698	2,081
Town05_Opt	600	420	120	60	880	883
Town07_Opt	600	420	120	60	901	2,342
Town10HD	600	420	120	60	1,405	1,027
Total	4,081	2,856	816	409	6,707	9,568

Weather Distribution

Weather	Frames
ClearNoon	692
ClearSunset	690
ClearNight	682
FoggyNoon	681
SnowNoon	669
DustStorm	667
Total	4,081

ClearNoon, ClearSunset, ClearNight, and DustStorm use CARLA weather presets. FoggyNoon and SnowNoon include custom rendering adjustments. SnowNoon simulates atmospheric haze and falling snow but does not model snow accumulation on surfaces.

Modalities

Every accepted frame has synchronized files for the following modalities:

Modality	Representation
Visible RGB	8-bit PNG, organized by weather
Camera depth	Dense float32 NPY in meters
Depth preview	16-bit PNG and color PNG
Surface normal	Camera-space float32 NPY and PNG preview
Target segmentation	8-bit class-ID PNG and color PNG
LiDAR	Raw float32 NPY point cloud in (x, y, z, intensity) format
Projected LiDAR	RGB-aligned sparse depth NPY plus 16-bit and color previews
Detection	YOLO TXT, COCO JSON, and per-frame JSON

Surface-normal channels use the camera convention x-right, y-down, and z-forward. Invalid pixels and depth-discontinuity pixels are stored as zero. Projected LiDAR values represent camera-forward depth in meters; zero means no LiDAR return at that pixel.

Detection Classes

YOLO class IDs are zero-based:

YOLO ID	Class
0	vehicle
1	pedestrian

Each YOLO label line follows:

class_id center_x center_y width height

All four coordinates are normalized to the image width and height. COCO category IDs are one-based: vehicle is 1 and pedestrian is 2.

The detection annotations are generated from CARLA vehicle and pedestrian actors, projected actor geometry, synchronized metric depth, and instance segmentation. Boxes represent filtered visible target pixels rather than an unfiltered full 3D projection.

Segmentation Values

The published target-semantic mask uses:

Pixel value	Meaning
0	Background
1	Trainable vehicle
2	Trainable pedestrian
255	Recognized target excluded from training

The color PNG is provided for visualization; use the ID PNG for training and evaluation.

Directory Structure

Each map is self-contained and its consolidated sequence directory has the same name as the map:

<MapName>/
|-- paired_weather/
|   `-- <MapName>/
|       |-- rgb/<Weather>/<frame>.png
|       |-- depth/
|       |   |-- npy/<frame>.npy
|       |   |-- vis_16bit/<frame>.png
|       |   |-- color/<frame>.png
|       |   `-- lidar/
|       |       |-- points/<frame>.npy
|       |       |-- projected_npy/<frame>.npy
|       |       |-- projected_vis_16bit/<frame>.png
|       |       `-- projected_color/<frame>.png
|       |-- surface_normal/
|       |   |-- npy/<frame>.npy
|       |   `-- png/<frame>.png
|       |-- segmentation/
|       |   |-- id/<frame>.png
|       |   `-- color/<frame>.png
|       |-- labels_yolo/<frame>.txt
|       |-- annotations/<frame>.json
|       |-- frame_index.csv
|       |-- groundtruth.txt
|       |-- groundtruth_multi.csv
|       `-- sequence_meta.json
|-- splits/{train,val,test}.txt
|-- splits_by_weather/
|-- coco/{train,val,test}.json
|-- data.yaml
|-- classes.txt
|-- dataset_manifest.json
|-- collection_config_snapshot.json
|-- quality_report.json
`-- QUALITY_REPORT.md

Paths stored in JSON, CSV, COCO, and split files are relative to the corresponding map directory.

Loading a Synchronized Frame

from pathlib import Path
import json
import numpy as np
from PIL import Image

dataset_root = Path("dataset_uav_multimap_town600_hutb300_ccsp300")
map_root = dataset_root / "Town02_Opt"
sequence_root = map_root / "paired_weather" / "Town02_Opt"

annotation = json.loads(
    (sequence_root / "annotations" / "000000.json").read_text(encoding="utf-8")
)

rgb = Image.open(map_root / annotation["image"]["rgb"])
depth_m = np.load(map_root / annotation["image"]["depth_npy_meters"])
normal = np.load(map_root / annotation["image"]["surface_normal_npy"])
segmentation = np.asarray(
    Image.open(map_root / annotation["image"]["segmentation"])
)

print(annotation["canonical_weather"])
print(rgb.size, depth_m.shape, normal.shape, segmentation.shape)

YOLO Usage

Each map includes a data.yaml. After downloading the complete repository, train from the selected map directory. For example:

yolo detect train model=yolo11x.pt data=Town02_Opt/data.yaml imgsz=1920 epochs=100

For a combined multi-map experiment, concatenate the paths from each map's split files into new root-level train, validation, and test lists. To evaluate cross-map generalization, keep one or more complete maps out of training rather than randomly mixing every map.

Quality Control

The collection pipeline applies synchronized-sensor checks, camera-view filters, minimum target-size and visibility rules, boundary-box rejection, weather balancing, and modality-file validation. All 4,081 published frames have matching RGB, depth, surface-normal, target-segmentation, YOLO, JSON, and LiDAR files.

Map-specific manifests, configuration snapshots, and quality reports are included so that the collection conditions can be inspected independently for each map.

Known Limitations

• This is synthetic data and has a simulation-to-real domain gap.
• Vehicle and pedestrian annotations cover CARLA actors. Baked static meshes without an actor identity are not guaranteed to receive detection labels.
• The custom CCSP map may contain occasional asset-streaming, mesh, or texture artifacts. Users should perform task-specific visual QA before training.
• Weather effects are simulation approximations and do not reproduce every real atmospheric or road-surface interaction.
• The supplied splits are frame-level splits within each map. They are not map-held-out domain-generalization splits.
• The repository license is marked as other; users must also comply with the licenses and usage terms of OpenHUTB/CARLA and the included map assets.

Intended Uses

The dataset is intended for research on UAV-view object detection, pedestrian and vehicle detection, small-object recognition, semantic segmentation, metric depth, surface-normal estimation, multimodal learning, robustness across weather, and cross-map generalization.

It is not intended for direct safety-critical deployment without validation on representative real-world data.

Citation

If this dataset contributes to a publication, cite the dataset repository and the OpenHUTB/CARLA simulator resources used to generate it. A paper-specific BibTeX entry can be added here when the accompanying publication is available.