Confidence-based Visual Dispersal for Few-shot Unsupervised Domain Adaptation (C-VisDiT)

Pytorch implementation of C-VisDiT (Confidence-based Visual Dispersal Transfer)

Overview

We present the Confidence-based Visual Dispersal Transfer learning method for Few-shot Unsupervised Domain Adaptation, aiming to comprehensively consider the importance of each sample during transfer based on its confidence.

Compared with state-of-the-art methods, PCS improves the mean classification accuracy over different domain pairs on FUDA by 4.4%/1.7% (1-shot/3-shots labeled source), 2.6%/2.8% (3%/6% labeled source), 1.5% (1% labeled source), and 2.0%/2.5% (1-shot/3-shots labeled source) on Office-31, Office-Home, VisDA-C, and DomainNet, respectively.

Requirements

conda create -n cvisdit python=3.7.11
conda activate cvisdit

conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
pip install dotmap faiss-gpu==1.7.0 scikit-learn tensorboard tqdm

Training

Download the datasets from the Internet (Split files are provided in data/splits, we now support Office, Office-Home, VisDA-2017, and DomainNet)
Soft-link the datasets under the data folder (alias for corresponding datasets are office, officehome, visda17, and domainnet)
To train the model, please refer to train.sh. We provide training configurations and SOTA result training-logs in the ./config and the ./logs folders. Please note that model performance is not sensitive to hyper parameters added by C-VisDiT (confidence_params). These hyper parameters can change in a wide range without greatly affecting the final results.

Citation

TODO

Acknowlegdement

This code is built on [PCS]. We thank the authors for sharing their code and some of the training configuration files. We reproduced some of the PCS results on our own.

ToDo List

Link to the official ICCV 2023 article page
Add bibtex citations
Link to the conference poster
Upload the article to ArXiv