- A new and complex traffic dataset for unstructured scenarios in India.
- Total dataset size is 100GB and increasing with more than 1000 one-minute video clips, over 2 million annotated frames with ego-vehicle trajectories, and more than 13 million bounding boxes.
- Up to 40 total agents and 9 unique agents per frame.
- Annotations include rare and interesting driving behaviors such as cut-ins, yielding, overtaking, overspeeding, zigzagging, and rule-breaking etc.
- Diverse traffic scenarios including rainy weather, nighttime driving, driving in rural areas with unmarked roads, and high-density traffic scenarios.
| Paper | Code | Dataset | |
|---|---|---|---|
| arXiv preprint | GitHub | Download Annotations and Raw Videos |
Research tasks
Download the pre-trained models for 2D object detection and action-behavior prediction from the GitHub link above.
2D object detection and tracking
State-of-the-art 2D object detection, which succeeds on existing datasets like Waymo, nuScenes, and Apolloscape, fails on our dataset due to the complexity and novelty of the object classes. The current best mAP on METEOR is 8.3. Improving this baseline calls for new techniques in computer vision and engineering.
Action-behavior prediction
We introduce a novel benchmark task called “action-behavior prediction”. Our goal is to predict both actions as well as rare and interesting cases. Action-behavior prediction is distinct from action prediction (which exclusively considers actions) and behavior prediction (which includes trajectory prediction). We provide a strong baseline with an mAP score of 70.74.
Trajectory forecasting
Another line of research on our dataset is trajectory forecasting as we also provide GPS coordinates of the ego-vehicle. There is a littany of work in trajectory or motion forecasting using deep recurrent neural networks and these can be used in conjunction with the RGB front-view camera views to perform motion forecasting.
Bibtex
Please cite our work if you found it useful,
@article{chandra2021meteor,
title={METEOR: A Massive Dense \& Heterogeneous Behavior Dataset for Autonomous Driving},
author={Chandra, Rohan and Mahajan, Mridul and Kala, Rahul and Palugulla, Rishitha and Naidu, Chandrababu and Jain, Alok and Manocha, Dinesh},
journal={arXiv preprint arXiv:2109.07648},
year={2021}
}
Acknowledgements
We thank NavAjna for creating this dataset. The dataset can alternately be found at their website.
Contact
Rohan Chandra (rchandr1@umd.edu).
License
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.