Skip to content

Latest commit

 

History

History
79 lines (65 loc) · 3.7 KB

README.md

File metadata and controls

79 lines (65 loc) · 3.7 KB
Raw

Learning CNN on ViT: A Hybrid Model to Explicitly Class-specific Boundaries for Domain Adaptation

This repository contains the code of the ECB method for Classification in Domain Adaptation.

Ba-Hung Ngo*, Nhat-Tuong Do-Tran*, Tuan-Ngoc Nguyen, Hae-Gon Jeon and Tae Jong Choi†
Accepted In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024).

PWC PWC PWC PWC PWC PWC

Proposed Method



  • Supervised Training: We train both ViT and CNN branches on labeled samples.
  • Finding To Conquering Strategy (FTC): We find class-specific boundaries based on the fixed ViT Encoder E1 by maximizing discrepancy between the Classifier F1 and F2. Subsequently, the CNN Encoder E2 clusters the target features based on those class-specific boundaris by minimizing discrepancy.

Prepare

Dataset

Please follow the instructions in DATASET.md to download datasets.

Installation

conda env create -f environment.yml

Training

  • The train.yaml is the config file for training our method. You can change the arguments to train Semi-Supervised Domain Adaptation (SSDA) or Unsupervised Domain Adaptation (UDA).
python train.py --cfg configs/train.yaml

Evaluation

  • If you need evaluate the test dataset with our pretrained model. You need to download these checkpoint.
sh download_pretrain.sh
  • For evaluation, you need to modify the configuration arguments in test/yaml in the configs folder. These arguments are described in CONFIG.md
python test.py --cfg configs/test.yaml

Visualization

  • The visualization compares features from two networks (CNN, ViT) for the real --> sketch on the DomainNet dataset in the 3-shot scenario, before and after adaptation with the FTC strategy.


  • The visualization in a few samples using GRAD-CAM technique to show to performance for CNN and ViT when applying ECB method.

Citation

@InProceedings{
    author    = {Ngo, Ba Hung and Do-Tran, Nhat-Tuong and Nguyen, Tuan-Ngoc and Jeon, Hae-Gon and Choi, Tae Jong},
    title     = {Learning CNN on ViT: A Hybrid Model to Explicitly Class-specific Boundaries for Domain Adaptation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {28545-28554}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.