Skip to content

Latest commit

 

History

History
73 lines (52 loc) · 2.27 KB

README.md

File metadata and controls

73 lines (52 loc) · 2.27 KB
Raw

Zero-VIRUS*: Zero-shot VehIcle Route Understanding System for Intelligent Transportation (CVPR 2020 AI City Challenge Track 1)

Authors: Lijun Yu, Qianyu Feng, Yijun Qian, Wenhe Liu, Alexander G. Hauptmann
Email: [email protected]

*Written in the era of Coronavirus Disease 2019 (COVID-19), with a sincere hope for a better world.

@inproceedings{yu2020zero,
  title={Zero-VIRUS: Zero-shot VehIcle Route Understanding System for Intelligent Transportation},
  author={Yu, Lijun and Feng, Qianyu and Qian, Yijun and Liu, Wenhe and Hauptmann, Alexander G.},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},
  year={2020}
}

Setup

Install miniconda, then create the environment and activate it via

conda env create -f environment.yml
conda activate zero_virus

Directory structure:

  • datasets
    • Dataset_A (AIC20_track1_vehicle_counting.zip/Dataset_A)
    • Dataset_B (hidden evaluation)
  • experiments
    • efficiency
      • aic2020-base.json
    • <experiment_name>
      • output.txt

Evaluate

As a zero-shot system, no training is required. We use Mask R-CNN pretrained on COCO from detectron2 as detector, whose weights will be downloaded automatically at the first run.

As the dataset only provided screenshots of the pre-defined routes, we created our own annotation of them with labelme.

To get system outputs, run

./evaluate.sh <experiment_name> <dataset_split>
# For example
./evaluate.sh submission Dataset_A

To get efficiency base score, run

python utils/efficiency_base.py

Performance

On Dataset A with 8 V100 GPUs:

  • S1: 0.9328
    • S1_Effectiveness: 0.9120
      • mwRMSE: 4.2738
    • S1_Efficiency: 0.9815
      • time: 3084.04
      • baseline: 0.546801

Visualizations available at Google Drive.

License

See LICENSE. Please read before use.