UrbanOmniView

A Multi-Perspective Dataset for Calibration-Free Monocular 3D Detection of Urban Traffic Participants

Mehmet Kerem Turkcan  Devika Gumaste  Zoran Kostic
AIDL Lab, Department of Electrical Engineering
Columbia University

UrbanOmniView is a dataset for calibration-free monocular 3D object detection across the camera viewpoints found in modern urban sensing: ego-vehicle dashcams, pole-mounted infrastructure cameras, and aerial drones. It combines real-world driving data, real-world infrastructure data, and high-fidelity synthetic data rendered in Unreal Engine 5.

Every object is annotated with a 2D bounding box, a class label, and eight ordered keypoints, the projections of its 3D bounding box corners onto the image plane. A single model trained on this format can perform both 2D detection and 3D reasoning without camera intrinsics at inference time. UrbanOmniView was introduced with UrbanOmniDetect at the CVPR 2026 DriveX workshop.

Dataset at a Glance

Source	Frames	Viewpoint	Provenance
KITTI	15,022	Ego-vehicle	Converted from the KITTI 3D object benchmark
DAIR-V2X	12,424	Infrastructure	Converted from the DAIR-V2X infrastructure split
UE5 Synthetic	10,000	Ground, infrastructure, drone	Rendered for this work and released here
Total	37,446

10% of UrbanOmniView is held out for testing.

Classes

Class	Description
car	Passenger vehicles, trucks, vans, and buses
person	Pedestrians
bike	Bicycles, motorcycles, and scooters

Annotation Format

Annotations follow the YOLO keypoint format. Each object line contains:

• Class label. An integer index into the class list above.
• Bounding box. Normalized center x, center y, width, and height.
• Keypoints. Eight ordered 2D points (x, y). Indices 0 to 3 are the top corners of the 3D bounding box, and indices 4 to 7 are the bottom corners that touch the ground plane.

The keypoint ordering is identical across all viewpoints, which lets a single model learn orientation-aware detection regardless of camera placement.

Data Sources

KITTI

The KITTI Vision Benchmark Suite provides images and 3D annotations from a car-mounted sensor rig in Karlsruhe, Germany. We use the left-camera images and project the 3D box labels into 2D keypoints with the provided calibration matrices. Calibration is used only to generate labels, never at inference time.

DAIR-V2X

DAIR-V2X is a vehicle-infrastructure cooperative dataset recorded at real intersections in Beijing. We use the infrastructure-side images, captured by elevated pole-mounted cameras looking down at traffic, and project the 3D annotations into 2D keypoints with the provided infrastructure camera parameters.

UE5 Synthetic

We rendered 10,000 frames in the Unreal Engine 5 City Sample. The generation pipeline provides:

• Dynamic environments. Weather and lighting variation, including rain, snow, and day and night cycles.
• Randomized traffic. Vehicle, pedestrian, and cyclist assets placed procedurally.
• Multi-viewpoint cameras. Ground-level, infrastructure-pole, and drone viewpoints sampled by a scripted camera rig.
• Ray-traced rendering. RGB output at 4K resolution with physically based lighting.
• Automatic annotation. 3D bounding boxes taken directly from engine object transforms and collision bounds, which yields pixel-accurate projected keypoints.

Usage

Download the dataset:

huggingface-cli download mehmetkeremturkcan/UrbanOmniView --repo-type dataset --local-dir .

Dataset configuration for Ultralytics YOLO:

path: ./urbanomniview/
train: '../.././urbanomniview_train.txt'
val: '../.././urbanomniview_val.txt'
test: '../.././urbanomniview_test.txt'

nc: 3
names: ['car', 'person', 'bike']
kpt_shape: [8, 2]

Train with the scripts from the GitHub repository:

python train.py --data cfg/dataset/urbanomniview.yaml

Or with Ultralytics directly:

from ultralytics import YOLO

model = YOLO("yolo11x-pose-p2.yaml").load("yolo11x.pt")
model.train(data="cfg/dataset/urbanomniview.yaml", imgsz=640, epochs=100)

Benchmark Results

Models trained on UrbanOmniView generalize across all three viewpoint categories. The best configuration, YOLO11x with the P2 feature level at 1920 × 1920, achieves:

Benchmark	Metric	Score
UrbanOmniView val	mAP50:95	0.808
KITTI	AP3D Moderate	30.71
KITTI	APBEV Moderate	35.19
DAIR-V2X val	AP @ OKS = 0.50	0.938

Calibration-dependent baselines trained on single-viewpoint data score near zero on viewpoints outside their training distribution. See the paper for detailed comparisons.

Ethical Considerations

• KITTI and DAIR-V2X contain real-world street imagery in which faces and license plates may be visible. Users should comply with the original dataset licenses and local privacy regulations.
• UE5 Synthetic data contains no real individuals.
• The dataset is intended for research in autonomous driving, traffic safety, and cooperative perception. We discourage its use for mass surveillance or any application that violates individual privacy.

Citation

@inproceedings{turkcan2026urbanomnidetect,
  title     = {Calibration-Free View-Agnostic Monocular {3D} Object Detection for Urban Scenes},
  author    = {Turkcan, Mehmet Kerem and Gumaste, Devika and Kostic, Zoran},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  pages     = {786--795},
  year      = {2026}
}

Acknowledgements

This work began while the first author was a postdoc in the Department of Electrical Engineering (AIDL Lab) at Columbia University. It was supported by the NSF Engineering Research Center for Smart Streetscapes under Award EEC-2133516, NSF Grants CNS-2450567 and CNS-2038984, and by computing resources from the NVIDIA Academic Grant Program and the Empire AI Consortium.

License

The UE5 synthetic portion is released under CC BY-NC 4.0. KITTI and DAIR-V2X remain subject to their original licenses.