Continual Test-Time Object Detection

This is the official code implementation for 'What, How, and When Should Object Detectors Update in Continually Changing Test Domains?' presented at CVPR 2024 arxiv.

Installation Instruction

We use Python 3.10, PyTorch 1.11.0 (CUDA 11.8 build). The codebase is built on Detectron2.

conda create -n cta_od python=3.10

Conda activate cta_od

conda install pytorch==1.11.0 torchvision==0.12.0 cudatoolkit=11.8 -c pytorch

cd ContinualTTA_ObjectDetection
pip install -r requirements.txt

## Make sure you have GCC and G++ version <=8.0
cd ..
python -m pip install -e ContinualTTA_ObjectDetection

Dataset Preparation

Please follow dataset structure below.

1. Coco & coco-corruption

   - coco
       - train2017
       - val2017
       - val2017-snow
       - val2017-frost
       ...
   

2. SHIFT

   - shift
       - discrete
           - images
               - train
                   - front
                       - images
                           ...
                       - det_2d.json
                       - seq.csv
               - val
                   - front
                       ...
       - continuous1x
       - continuous10x
   

Model Checkpoints and Feature Statistics

You can download the following assets via Link.

• Source Model checkpoints
• Feature Statistics
• Dataset (COCO-Corrupted, 15 corruption type)

Run Test-Time Adaptation

# coco -> coco-corrupted
bash scripts/coco_adapt.sh

# shift discrete
bash scripts/shift_discrete_adapt.sh

# shift continuous
bash scripts/shift_continuous_adapt.sh

Citation

If you found our work is useful in your research, please consider starring ⭐ us on GitHub and citing 📚 us in your research!

@InProceedings{Yoo_2024_CVPR,
    author    = {Yoo, Jayeon and Lee, Dongkwan and Chung, Inseop and Kim, Donghyun and Kwak, Nojun},
    title     = {What How and When Should Object Detectors Update in Continually Changing Test Domains?},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {23354-23363}
}