Skip to content

Latest commit

 

History

History
112 lines (83 loc) · 4.19 KB

README.md

File metadata and controls

112 lines (83 loc) · 4.19 KB
Raw

About

This is the source code of the paper Assessing Domain Gap for Continual Domain Adaptation in Object Detection

If you find our code is useful, please cite:

@article{Doan23DGECDA,
    author    = {Anh-Dzung Doan and Bach Long Nguyen and Surabhi Gupta and Ian Reid and Markus Wagner and Tat-Jun Chin},
    title     = {Assessing Domain Gap for Continual Domain Adaptation in Object Detection},
    journal   = {Computer Vision and Image Understanding},
    year      = {2024}
}

Installation

Our installation is mainly based on Detectron2's installation with few modifications

  • Clone this repo, suppose the source code is saved in [ROOT_DIR]/DGE-CDA/

  • Install Python >= 3.7 (tested on Python 3.9)

  • Install Pytorch >= 1.8 and torchvision that matches the Pytorch installation (tested on Pytorch 1.11, torchvision 0.12, CUDA 11.3)

  • Install OpenCV

    pip install opencv-python

    (tested on opencv-python 4.5.5.64)

  • Install detectron2

    cd [ROOT_DIR] 
python -m pip install -e DGE-CDA 
cd [ROOT_DIR]/DGE-CDA

  • Install POT library

    pip install POT

  • Install pyJoules

    pip install pyJoules[nvidia]

  • Install tqdm

    pip install tqdm

Dataset preparation

  • Download the dataset DGTA_SeaDronesSee_jpg from DeepGTAV (suppose save it to [ROOT_DIR]/DGTA_SeaDronesSee_jpg)

  • Create folder DGTA_SeaDronesSee_merged/ inside [ROOT_DIR]/DGE-CDA/datasets/

  • Create folder images/ inside [ROOT_DIR]/DGE-CDA/datasets/DGTA_SeaDronesSee_merged/

  • Copy all images from [ROOT_DIR]/DGTA_SeaDronesSee_jpg/train/images/ to [ROOT_DIR]/DGE-CDA/datasets/DGTA_SeaDronesSee_merged/images/

  • Copy all images from [ROOT_DIR]/DGTA_SeaDronesSee_jpg/val/images/ to [ROOT_DIR]/DGE-CDA/datasets/DGTA_SeaDronesSee_merged/images/

  • Create folder experiments/ inside [ROOT_DIR]/DGE-CDA/datasets/DGTA_SeaDronesSee_merged/

  • Download all json files from HERE and put them in [ROOT_DIR]/DGE-CDA/datasets/DGTA_SeaDronesSee_merged/experiments/

Finally, we should obtain

  • all images located in [ROOT_DIR]/DGE-CDA/datasets/DGTA_SeaDronesSee_merged/images/

  • and all json files located in [ROOT_DIR]/DGE-CDA/datasets/DGTA_SeaDronesSee_merged/experiments/

DEMO

  • Pretrain RetinaNet on source domain

    python cda/pretrain_source.py --config-file configs/dgta.yaml

    After the training finishes, we should find model_final.pth located in work_dir/DeepGTAV/CLEAR_9-15/

  • Randomly generate random projections

    python cda/generate_random_projections.py --save-dir work_dir/DeepGTAV/projections

    We should find 3 projections in the folder work_dir/DeepGTAV/projections

  • To plot the histograms of AP discrepancy and domain gaps (Fig.5 in the paper), run

    python cda/plot_correlation.py --config-file configs/dgta.yaml --projs-dir work_dir/DeepGTAV/projections

    This script will take about 4h in Intel-Core-i7 with 16GB RAM and NVIDIA GeForce GTX 1080Ti

  • To reproduce the results in Table 5,

    • For SDA with DGE
    python cda/cont_dgta.py --config-file configs/dgta_cont.yaml --projs-dir work_dir/DeepGTAV/projections --gap-thr 0.02 --dge

    The trained models can be found in work_dir/DeepGTAV/adapt_with_dge/. The energy consumption record can be found in its subfolders, e.g., work_dir/DeepGTAV/adapt_with_dge/CLEAR_9-15_and_OVERCAST_0-1/energy.csv

    • For SDA w/o DGE
    python cda/cont_dgta.py --config-file configs/dgta_cont.yaml

    The trained models can be found in work_dir/DeepGTAV/adapt_wo_dge/. The energy consumption record can be found in its subfolders.

    • It should be noted that
      • The energy unit is Joule
      • there is a common error Permission denied: '/sys/class/powercap/intel-rapl/intel-rapl:0/energy_uj'. To resolve it, simply grant the permission
      sudo chmod -R 7777 /sys/class/powercap/intel-rapl/intel-rapl:0