ArXiv | Models

Models trained on UVH-26 deliver up to 31.5% higher mAP than COCO-pretrained baselines, demonstrating significant gains in real-world performance for Indian traffic scenarios.

Dataset Card for UVH-26 (Urban Vision Hackathon Dataset)

Dataset Summary

UVH-26 is a large-scale, India-specific traffic-camera image dataset released by AIM @ IISc for research in intelligent transportation systems and vehicle detection.
It contains 26,646 high-resolution (1080p) frames sampled from ≈ 2,800 Bengaluru Safe City CCTV cameras over a 4-week period.
Images were annotated through a nationwide crowdsourced hackathon involving 565 college students, producing ≈ 1.8 million bounding boxes across 14 fine-grained vehicle classes representative of Indian traffic conditions.

To capture different levels of annotation consensus, UVH-26 includes two separate annotation sets:

1. UVH-26-MV — final labels computed via majority voting across multiple annotators per image.
2. UVH-26-ST — labels generated using the STAPLE algorithm (an Expectation–Maximization–based probabilistic consensus method) for higher reliability.

These versions share identical image data but differ in bounding box consensus logic.

Attribution

More technical details about the dataset and models are available in our Technical Report available on arXiv. If you use these datasets or models, kindly cite the following: The Urban Vision Hackathon Dataset and Models: Towards Image Annotations and Accurate Vision Models for Indian Traffic, Preliminary Dataset Release, UVH-26-v1.0, Akash Sharma, Chinmay Mhatre, Sankalp Gawali, Ruthvik Bokkasam, Brij Kishore, Vishwajeet Pattanaik, Tarun Rambha, Abdul R. Pinjari, Vijay Kovvali, Anirban Chakraborty, Punit Rathore, Raghu Krishnapuram and Yogesh Simmhan, Technical Report, Indian Institute of Science, arXiv.2511.02563, Nov, 2025.

@techreport{sharma2025uvh26,
  title        = {Towards Image Annotations and Accurate Vision Models for Indian Traffic, Preliminary Dataset Release, UVH-26-v1.0},
  author       = {Akash Sharma and Chinmay Mhatre and Sankalp Gawali and Ruthvik Bokkasam and Brij Kishore and Vishwajeet Pattanaik and Tarun Rambha and Abdul R. Pinjari and Vijay Kovvali and Anirban Chakraborty and Punit Rathore and Raghu Krishnapuram and Yogesh Simmhan},
  institution  = {Indian Institute of Science},
  type         = {Technical Report},
  number       = {arXiv:2511.02563},
  year         = {2025},
  month        = {November},
  doi          = {10.48550/arXiv.2511.02563}
}

Dataset Structure

The datasets released follow the folder structure described below.

**1. UVH-26-Train/**

Contains 80% of the UVH-26 dataset used for training.

• images/ – Training images organized into subfolders (000/, 001/, …) for convenience.
  • images/000/* – Actual training images (1.png, 2.png, …). Each image filename is unique across the entire dataset.
  • images/001/*, etc. – Additional subfolders following the same structure.
• UVH-26-MV-Train.json – Majority Voting consensus annotations for training images in COCO JSON format.
• UVH-26-ST-Train.json – STAPLE consensus annotations for training images in COCO JSON format.

**2. UVH-26-Val/**

Contains 20% of the UVH-26 dataset used for validation.

• images/ – Validation images organized into subfolders (000/, 001/, …).
  • images/000/* – Actual validation images. All filenames are globally unique across both training and validation sets.
  • images/001/*, etc. – Additional subfolders following the same structure.
• UVH-26-MV-Val.json – Majority Voting consensus annotations for validation images in COCO JSON format.
• UVH-26-ST-Val.json – STAPLE consensus annotations for validation images in COCO JSON format.

Annotation JSON Schema

Each annotation file follows the standard COCO structure:

• images — list of image metadata
  id, file_name, width, height
• annotations — object instances
  id, image_id, category_id, bbox [x, y, width, height], area
• categories — class taxonomy (IDs and names below)

Annotation Pipeline

• Source: frames captured between 06:00 – 18:00 IST during February 2025
• Pre-annotation: generated using a fine-tuned RT-DETR v2-X model trained on ≈ 3 k expert-labeled images
• Crowdsourcing: > 550 student volunteers corrected or validated predictions through a gamified web interface with leaderboards
• Consensus: both majority voting and STAPLE algorithms applied to derive final annotations

Vehicle Classes

Collection and Processing

• Source: ≈ 2,800 Safe City surveillance cameras operated by Bengaluru Police
• Coverage: both junction and mid-block perspectives across multiple city zones
• Selection: images with high vehicle density, occlusion, and diverse viewpoints prioritized

Intended Uses

• Building accurate, lightweight, edge-deployed perception systems for Intelligent Transportation Systems (ITS)
• Training and benchmarking vehicle detection models

License

• Dataset: CC BY 4.0 International
• Pre-trained Models: Apache 2.0

Acknowledgements

We thank the Bengaluru Traffic Police (BTP) and the Bengaluru Police for providing access to the Safe City camera data from which the image datasets used for this release were derived.
We thank Capital One for sponsoring the prizes for the Urban Vision Hackathon competition.
We thank IISc’s AI and Robotics Technology Park (ARTPARK) and the Centre for Infrastructure, Sustainable Transportation and Urban Planning (CiSTUP) for funding the annotation and model-training efforts, and the Kotak IISc AI-ML Centre (KIAC) for providing the GPU resources required to train the models.
We acknowledge the outreach support provided by the ACM India Council and the IEEE India Council to encourage chapter volunteers to participate in the hackathon.
Lastly, we thank the AI Centers of Excellence (AI COE) initiative of the Ministry of Education, their Apex Committee members, and the AIRAWAT Research Foundation, whose support helped catalyze these efforts.

Created by the AI for Integrated Mobility (AIM) group at the Indian Institute of Science (IISc), Bengaluru.