Set Up Environment

Our code is built on top of HTCN. We follow the paper in our implementation details and Faster-RCNN respository to setup the environment. We use python 3.8, pytorch 1.7.1 and a single 3090 for training. You can set up your own environment according to your GPU computing power and you can install other packages by running 'pip install -r requirements.txt' (this should be done when the environment is activated and Faster-RCNN has already been set up.) You can run:'python setup.py build develop' to set up faster-rcnn. If you have problems setting up, you may find solutions in issues of HTCN or Faster-RCNN respository and use provided setting up files in lib/build. You may need this command: export TORCH_CUDA_ARCH_LIST="8.0" to help you make your GPU compatible with your CUDA version.

 

Dataset Preparation

For double blind principle, we have removed all external links. We will add the links of datasets and models in our open-source version to help readers implement our code more easily. Please download the dataset from official dataset websites. All codes are written to fit for the format of PASCAL_VOC. Please prepare the annotations in VOC format.
Please download vgg16_caffe.pth and change the path in lib/model/utils/config.py and change the data path in lib/datasets/config_dataset.py. Also change your path in lib/model/utils/parser_func.py in line 416 to change cfg_path. After preparation ,your dataset should looks like:
|--cityscape_t
 |--Annotations(VOC format)
 |--ImageSets
   |--Main
    |--train.txt
    |--val.txt
    |--trainval.txt
 |--JPEGImages

 

Training scripts

We take cityscapes -> foggy-cityscapes as an example. We have 2975 images for training and 500 images for test for both cityscapes and foggy-cityscapes, respectively. We name the folders of these two datasets as cityscapes_s and cityscapes_t. This is a four stage training process:

Stage1: Pretrain model using cityscapes training dataset

python trainval_pretrain_adv.py --dataset cs --net vgg16 --log_ckpt_name "your own path" --save_dir "your own path"

Stage2: Select source_similar data and source_dissimilar data

1. Duplicate the folder cityscape_t to two folders cityscapes_t_similar and cityscapes_t_disimilar

2. Change line 224 and 231 in select_by_uncertainty.py to localize the path of train.txt in foggy_cityscapes_similar(disimilar)/ImageSets/Main

python select_by_uncertainty.py --dataset_t cs_fg --net vgg16 --log_ckpt_name "your own path" --save_dir "your own path" --load_name "path of checkpoint of stage1"

Stage3: TSD adversarial training with mean teacher

python trainval_adv_mt.py --epochs 30 --dataset cs_fg_similar --dataset_t cs_fg_disimilar --net vgg16 --log_ckpt_name "your own path" --save_dir "your own path" --load_name "path of checkpoint of stage1"

Stage4: Fine-tune with False negative simulation

python trainval_adv_mosaic_mt.py --epochs 20 --dataset cs_fg --dataset_t cs_fg_similar --net vgg16 --log_ckpt_name "your own path" --save_dir "your own path" --load_name "path of checkpoint of stage3"

 

Test

python test_net_adv.py --dataset_t cs_fg --net vgg16 --log_ckpt_name "your own path" --save_dir "your own path" --load_name "path of checkpoint of stage4"