buaali / da-faster-rcnn-pytorch

An unofficial implementation of 'Domain Adaptive Faster R-CNN for Object Detection in the Wild ’

da-faster-rcnn-PyTorch

A PyTorch implementation of 'Domain Adaptive Faster R-CNN for Object Detection in the Wild ’

Preparation

Requirements: Python=3.6 and Pytorch=0.4.0

1. Install Pytorch

2. Our code is conducted based on faster-rcnn.pytorch,please setup the framework by it.

3. Download dataset

   • we use cityscape and cityscapes-foggy datasets respectly as source and target,the cityscapes dataset could be download Here

   • the format of datasets is similar with VOC,you just need to split train.txt to train_s.txt and train_t.txt

   • you can also download the dataset GoogleDrive

Train and Test

1.train the model,you need to download the pretrained model [vgg_caffe](https://github.com/jwyang/faster-rcnn.pytorch） which is different with pure pytorch pretrained model

2.change the dataset root path in ./lib/model/utils/config.py and some dataset dir path in ./lib/datasets/cityscape.py,the default data path is ./data

3 Train the model

# train cityscapes -> cityscapes-foggy
CUDA_VISIBLE_DEVICES=GPU_ID python da_trainval_net.py --dataset cityscape --net vgg16 --bs 1 --lr 2e-3 --lr_decay_step 6 --cuda

# Test model in target domain 
CUDA_VISIBLE_DEVICES=GPU_ID python eval/test.py --dataset cityscape --part test_t --model_dir=# The path of your pth model --cuda

Our model could arrive mAP=30.71% in target domain which is high than baseline mAP=24.26%