pinglmlcv / ROS

Rotation-based Open Set

ROS (Rotation-based Open Set)

PyTorch official implementation of "On the Effectiveness of Image Rotation for Open Set Domain Adaptation" in European Conference on Computer Vision 2020, ECCV2020

Experiments

In order to replicate the results shown in the paper (Tables 1,2) please follow these instructions:

1. Download Office-31 and Office-Home datasets:

   • Office-31: https://drive.google.com/file/d/0B4IapRTv9pJ1WGZVd1VDMmhwdlE/view save the folder as "office" in ROS folder. At the end you should have: office/amazon, office/webcam, office/dslr

   • Office-Home: https://drive.google.com/file/d/0B81rNlvomiwed0V1YUxQdC1uOTg/view save the folder as "office-home" in ROS folder. At the end you should have: office-home/OfficeHomeDataset_10072016/art, office-home/OfficeHomeDataset_10072016/clipart, office-home/OfficeHomeDataset_10072016/real_world, office-home/OfficeHomeDataset_10072016/product

2. Use the python version: Python 3.6.8

   Install all the libreries requested with the command:

   pip3 install -r requirements_ROS.txt

   Please note that, for the sake of convenience, we also provide a Dockerfile to directly create a docker container with all the necessary requirements.

3. Go into the folder ROS and:

   3a. In order to replicate the experiments of Office31 dataset with ResNet-50 (Table 1) run:

    train_resnet50_office31.sh replacing 
    "/.../" with "/path_in_which_you_save_ROS/"
   

   3b. In order to replicate the experiments of Office-Home dataset with ResNet-50 (in Table 2) run:

    train_resnet50_officehome.sh replacing 
    "/.../" with "/path_in_which_you_save_ROS/"        
   

   3c. In order to replicate the experiments of Office31 dataset with VggNet (in Table 1) run:

    train_vgg_office31.sh replacing 
    "/.../" with "/path_in_which_you_save_ROS/"
   

You can also replicate the results obtained for STA_max,STA_sum,OSBP and UAN (Tables 1,2) following the instructions of the GitHub repositories proposed by the authors:

• STA: https://github.com/thuml/Separate_to_Adapt

• OSBP: https://github.com/ksaito-ut/OPDA_BP

• UAN: https://github.com/thuml/Universal-Domain-Adaptation

Citation

To cite, please use the following reference:

@inproceedings{BucciLoghmaniTommasi2020,
  title={On the Effectiveness of Image Rotation for Open Set Domain Adaptation},
  author={Silvia Bucci, Mohammad Reza Loghmani, Tatiana Tommasi},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2020}
}