Cross-domain Detection via Graph-induced Prototype Alignment

Introduction

This project is an implementation of ``Cross-domain Detection via Graph-induced Prototype Alignment'' in PyTorch, which is accepted by CVPR 2020. We would like to appreciate for the excellent work of jwyang/faster-rcnn.pytorch which lays a solid foundation for our work.

More details of this work can be found in our paper: [Paper (arxiv)].

Preparation

First of all, clone the code

git clone https://github.com/ChrisAllenMing/GPA-detection.git

Then, create a folder:

cd GPA-detection && mkdir data

Prerequisites

Python 2.7 or 3.6
Pytorch 0.4.1
CUDA 8.0 or higher

Dataset Preparation

Cityscapes & Foggy Cityscapes

p.s. Transform original segmentation annotations to detection formats using these codes.
SIM 10k
KITTI

Pretrained Model

In all experiments, we use a ResNet-50 model pretrained on ImageNet as the backbone. You can download this model from:

ResNet-50: Google Drive

This model should be put into ./data/pretrained_model/.

NOTE. We experimentally found that the Caffe pretrained model outperforms the PyTorch pretrained one. If you would like to evaluate our method with other backbones, a converted model from Caffe to PyTorch maybe favored.

Compilation

Install all the python dependencies using pip:

pip install -r requirements.txt

Compile the cuda depended modules, e.g. NMS, ROI Pooling, ROI Align and ROI Crop, using following simple commands:

cd lib
sh make.sh

Some issues about compilation can be found here.

Training

To train the Faster R-CNN baseline, simply run:

CUDA_VISIBLE_DEVICES=$GPU_ID python train_baseline.py --dataset sim10k \
                    --model_config baseline --net res50 --bs 3 --nw 1 \
                    --epochs 10 --lr 0.001 --lr_decay_step 6 \
                    --lr_decay_gamma 0.1 --cuda

To train the proposed model without graph-based aggregation, simply run:

CUDA_VISIBLE_DEVICES=$GPU_ID python train_GPA.py --dataset sim10k \
                    --tgt_dataset city --model_config adapt --mode adapt \
                    --rpn_mode adapt --net res50 --bs 3 --nw 1 --epochs 10 \
                    --lr 0.001 --lr_decay_step 6 --lr_decay_gamma 0.1 \
                    --warm_up 200 --da_weight 1.0 --rpn_da_weight 1.0 --cuda

To train the proposed model with graph-based aggregation, simply run:

CUDA_VISIBLE_DEVICES=$GPU_ID python train_GPA.py --dataset sim10k \
                    --tgt_dataset city --model_config gcn_adapt \
                    --mode gcn_adapt --rpn_mode gcn_adapt --net res50 --bs 3 \
                    --nw 1 --epochs 10 --lr 0.001 --lr_decay_step 6 \
                    --lr_decay_gamma 0.1 --warm_up 200 --da_weight 1.0 \
                    --rpn_da_weight 1.0 --cuda

In our experiments, we train the model with two GPUs as follows:

CUDA_VISIBLE_DEVICES=$GPU_IDs python train_GPA.py --dataset sim10k \
                    --tgt_dataset city --model_config gcn_adapt_multi \
                    --mode gcn_adapt --rpn_mode gcn_adapt --net res50 --bs 6 \
                    --nw 2 --epochs 10 --lr 0.001 --lr_decay_step 6 \
                    --lr_decay_gamma 0.1 --warm_up 200 --da_weight 1.0 \
                    --rpn_da_weight 1.0 --cuda --mGPUs

Test

In order to test the model of every epoch, for the baseline, you can run:

python iterative_test.py --dataset city --model_config baseline --net res50 \
                        --checksession $session_num --checkpoint $iter_num \
                        --start_epoch 1 --end_epoch 10 --gpu_id $GPU_ID \
                        --test_mode baseline --cuda

To evaluate the proposed GPA model, you can run:

python iterative_test.py --dataset city --model_config gcn_adapt --net res50 \
                        --checksession $session_num --checkpoint $iter_num \
                        --start_epoch 1 --end_epoch 10 --gpu_id $GPU_ID \
                        --test_mode GPA --cuda

Citation

If this work helps your research, please cite the following paper.

inproceedings{xu2020cross-domain,
  author={Minghao Xu and Hang Wang and Bingbing Ni and Qi Tian and Wenjun Zhang},
  title={Cross-domain Detection via Graph-induced Prototype Alignment},
  booktitle={{IEEE} Conference on Computer Vision and Pattern Recognition},
  pages={12355-12364},
  year={2020}
}