Learning Transformations to reduce Geometric Shift in Object Detection

[ Paper ]

Installation

This code is based on detectron2 and requires python>= 3.6

pip install -r requirements.txt

Dataset

Set the environ variable DETECTRON2_DATASETS to the parent folder of the datasets

/datasets
    /cityscapes
    /kitti
    /mot

Download

1. Cityscapes & Kitti from -- https://github.com/chengchunhsu/EveryPixelMatters#dataset

2. MOT sequence MOT20-02 -- https://motchallenge.net

Base Training

python train_net.py --config-file configs/<file_name>_random_crop.yaml 
.py

Aggregator Training

python train_net_only_aggregator.py --config-file configs/<file_name>_aggregator_fivestnperspective.yaml 
.py

Mean Teacher Training

We train on single V100 GPU with batch size 2 for mean teacher(in config setting which means 2 source domain and 2 target domain), steps

python train_net_student_teacher_<task>.py --config-file configs/<task>_student_teacher.yaml SOLVER.BASE_LR 1e-3 SOLVER.STEPS [10000,] MODEL.STN_ARCH FIVE_OPT_PERSPECTIVE
.py

Citation

@inproceedings{vidit2023learning,
  title={Learning Transformations To Reduce the Geometric Shift in Object Detection},
  author={Vidit, Vidit and Engilberge, Martin and Salzmann, Mathieu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={17441--17450},
  year={2023}
}