pretty_name: VisionDrive3D
license: mit
task_categories:
- object-detection
- image-segmentation
- depth-estimation
tags:
- synthetic
- autonomous-driving
- 3d-object-detection
- kitti
- coco
- yolo
- opengl
size_categories:
- 1K<n<10K
VisionDrive3D: Synthetic Driving Dataset
VisionDrive3D is a synthetic urban driving dataset rendered with a custom OpenGL engine. Every scene comes with an RGB image, a dense metric depth map, a segmentation mask, and annotations in COCO, YOLO, and KITTI (3D) formats. It also includes per-scene camera parameters (intrinsics and extrinsics) and object metadata.
- Code & project page: https://github.com/BenBunBo/VisionDrive3D
- File:
VisonDrive3D_testdata.zip(~1.1 GB compressed, ~6.6 GB extracted)
Summary
| Property | Value |
|---|---|
| Scenes | 1,000 |
| Resolution | 1600 × 900 |
| Annotated class | car (2,585 instances) |
| Splits | train 800 / val 100 / test 100 |
| Modalities | RGB, depth (float32, metres), segmentation mask |
| Label formats | COCO JSON, YOLO TXT, KITTI 3D TXT |
| Camera | Pinhole, fx = fy = 1200, cx = 800, cy = 450, FOV 45°, near 0.1, far 150 |
Installation
1. Download
Option A: hf CLI (recommended)
pip install -U huggingface_hub
hf download BenBunBo/VisionDrive3D VisonDrive3D_testdata.zip \
--repo-type dataset --local-dir ./data
Option B: Python
from huggingface_hub import hf_hub_download
zip_path = hf_hub_download(
repo_id="BenBunBo/VisionDrive3D",
filename="VisonDrive3D_testdata.zip",
repo_type="dataset",
local_dir="./data",
)
print(zip_path)
Option C: Direct link / git
wget https://huggingface.co/datasets/BenBunBo/VisionDrive3D/resolve/main/VisonDrive3D_testdata.zip
# or
git lfs install
git clone https://huggingface.co/datasets/BenBunBo/VisionDrive3D
2. Extract
The archive has no top-level folder, so extract it into its own directory:
mkdir -p data/output_dataset
unzip data/VisonDrive3D_testdata.zip -d data/output_dataset
import zipfile
zipfile.ZipFile("data/VisonDrive3D_testdata.zip").extractall("data/output_dataset")
3. Verify
ls data/output_dataset
# depth images ImageSets labels masks metadata README.txt
ls data/output_dataset/images | wc -l # 1000
wc -l data/output_dataset/ImageSets/*.txt # 800 / 100 / 100
4. Use with the VisionDrive3D code (optional)
git clone https://github.com/BenBunBo/VisionDrive3D
cd VisionDrive3D
# Put the extracted folder at ./output_dataset (the default path used by the scripts)
ln -s /path/to/data/output_dataset ./output_dataset
python src/ai_demo/run_all.py --dataset ./output_dataset
Folder structure
output_dataset/
├── README.txt
├── images/ # RGB renders scene_XXXXX.png (1600×900, 8-bit RGB)
├── depth/ # Depth maps scene_XXXXX.npy (900×1600, float32, metres)
├── masks/ # Segmentation masks scene_XXXXX.png (1600×900, 8-bit RGB colour-coded)
├── labels/
│ ├── coco/ # COCO JSON scene_XXXXX.json
│ ├── yolo/ # YOLO TXT scene_XXXXX.txt
│ └── kitti/ # KITTI 3D TXT scene_XXXXX.txt
├── metadata/
│ ├── scene_XXXXX.json # Camera + per-object metadata
│ └── dataset_log.csv # One row per scene (camera position, depth range, #objects, ...)
└── ImageSets/
├── train.txt # 800 scene ids
├── val.txt # 100 scene ids
└── test.txt # 100 scene ids
Scenes are named scene_00000 … scene_00999. Each file in ImageSets/ lists one
5-digit id per line (e.g. 00042 → scene_00042).
Annotation formats
YOLO (labels/yolo/*.txt): one object per line, normalised to [0, 1]:
<class_id> <x_center> <y_center> <width> <height>
0 0.976250 0.516667 0.047500 0.044444
Class 0 = car.
COCO (labels/coco/*.json): one JSON file per scene with images, annotations
(bbox as [x, y, w, h] in pixels, polygon segmentation, area) and categories.
The category list contains car (0), traffic_light (1), traffic_sign (2),
pedestrian (3) and entity (99), but only car is annotated in this release.
Note: images[].file_name has the output_dataset/ prefix, e.g.
output_dataset/images/scene_00000.png.
KITTI (labels/kitti/*.txt): standard KITTI object label format, 15 fields per line:
type truncated occluded alpha x1 y1 x2 y2 h w l x y z rotation_y
Car 1.00 0 2.15 1524 445 1600 485 1.88 2.16 4.88 37.32 1.66 56.71 1.57
Dimensions and location are in metres in camera coordinates. The camera intrinsics are
in metadata/scene_XXXXX.json → camera.intrinsic_matrix.
Depth (depth/*.npy): float32 array of shape (900, 1600) with per-pixel depth
in metres, clipped to the far plane (150 m).
Metadata (metadata/scene_XXXXX.json): scene_id, timestamp, image_path,
camera (position, target, up, fov_deg, near, far, intrinsic_matrix,
extrinsic_matrix), objects (class, world pose, scale, 2D/3D boxes, visibility,
occlusion ratio), render_config, and stats.
Quick start
import json
import numpy as np
from PIL import Image
root = "data/output_dataset"
sid = "scene_00000"
rgb = np.array(Image.open(f"{root}/images/{sid}.png")) # (900, 1600, 3) uint8
depth = np.load(f"{root}/depth/{sid}.npy") # (900, 1600) float32, metres
mask = np.array(Image.open(f"{root}/masks/{sid}.png")) # (900, 1600, 3) uint8
meta = json.load(open(f"{root}/metadata/{sid}.json"))
K = np.array(meta["camera"]["intrinsic_matrix"]) # 3×3
with open(f"{root}/labels/yolo/{sid}.txt") as f:
boxes = [list(map(float, line.split())) for line in f]
train_ids = open(f"{root}/ImageSets/train.txt").read().split()
print(rgb.shape, depth.min(), depth.max(), len(boxes), len(train_ids))
Train YOLO (Ultralytics): write a data.yaml that points at the split lists:
from pathlib import Path
root = Path("data/output_dataset").resolve()
for split in ["train", "val", "test"]:
ids = (root / "ImageSets" / f"{split}.txt").read_text().split()
(root / f"{split}_images.txt").write_text(
"\n".join(str(root / "images" / f"scene_{i}.png") for i in ids)
)
YOLO looks for labels next to the images (images/ → labels/), so copy or symlink
labels/yolo/*.txt into a sibling labels/ folder of the image directory, or use the
conversion scripts in the GitHub repo.
Citation
@misc{visiondrive3d2026,
title = {VisionDrive3D: Synthetic Driving Data for Detection and Segmentation},
author = {BenBunBo},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/BenBunBo/VisionDrive3D}}
}