Improving Transferability for Domain Adaptive Detection Transformers

Kaixiong Gong, Shuang Li, Shugang Li, Rui Zhang, Chi Harold Liu, Qiang Chen

ACM Multimedia 2022, Lisbon, Portugal.

[Supp]

This repository contains the code of our ACM MM 2022 work "Improving Transferability for Domain Adaptive Detection Transformers".

If you found our paper or this project is useful to you, please cite:

@article{gong2022improving,
  title={Improving Transferability for Domain Adaptive Detection Transformers},
  author={Gong, Kaixiong and Li, Shuang and Li, Shugang and Zhang, Rui and Liu, Chi Harold and Chen, Qiang},
  journal={arXiv preprint arXiv:2204.14195},
  year={2022}
}

Getting Started

Prepare data

The second step is to prepare datasets. Four datasets are involved:

• Cityscapes [download] (There are a lot of versions of Cityscapes, make sure you download the right one)
• Foggy Cityscapes [downlaod]
• BDD100k [download]
• Sim10k [download]

Processing dataset annotations. All annotations are should in COCO format which can be processed by the code. We should covert the above four datasets in to CoCo Format.

• Cityscapes to CoCo Format: using this script. (You can also refer to this project)
• BDD100k to CoCo Format: using this script.
• Sim10k to CoCo Format: using this script.

We also prived processed data lists in CoCo Format.

• Cityscapes and Foggy Cityscapes default lists
• Cityscapes car only lists (for Sim10k to Cityscapes)
• BDD100k data lists
• Sim10k data list

Requirements

See Deformable DETR for more details.

• Linux, CUDA>=9.2, GCC>=5.4

• Python>=3.7

  We recommend you to use Anaconda to create a conda environment:

  conda create -n deformable_detr python=3.7 pip

  Then, activate the environment:

  conda activate deformable_detr

• PyTorch>=1.5.1, torchvision>=0.6.1 (following instructions here)

  For example, if your CUDA version is 9.2, you could install pytorch and torchvision as following:

  conda install pytorch=1.5.1 torchvision=0.6.1 cudatoolkit=9.2 -c pytorch

• Other requirements

  pip install -r requirements.txt

Compiling CUDA operators

cd ./models/ops
sh ./make.sh
# unit test (should see all checking is True)
python test.py

Training

To train our method O2net on Cityscapes to Foggy Cityscapes, first pre-train a source model:

CUDA_VISIBLE_DEVICES=* GPUS_PER_NODE=n ./tools/run_dist_launch.sh n ./configs/r50_deformable_detr.sh --output_dir exps/source_model --dataset_file city2foggy_source

then:

CUDA_VISIBLE_DEVICES=* GPUS_PER_NODE=n ./tools/run_dist_launch.sh n ./configs/DA_r50_deformable_detr.sh --output_dir exps/ours --transform make_da_transforms --dataset_file city2foggy --checkpoint exps/source_model/checkpoint.pth

Or simply run:

sh DA.sh

Testing

Main results

Model	Source Domain	Target Domain	mAP@50	Download
Source_only	Cityscapes	Foggy Cityscapes	28.6	Model
O2Net	Cityscapes	Foggy Cityscapes	47.2	Model

Note: The batch size is set as 2 on a single GPU.

Acknowledgement

We thank the contributors of Deformable DETR for their great work. We build our method on it.