Impact-of-ImageNet-Model-Selection-on-Domain-Adaptation

This directory contains the code for paper Impact of ImageNet Model Selection on Domain Adaptation, which is accepted by In 2020 IEEE Winter Applications of Computer Vision Workshops (WACVW).

And the code for paper Modified Distribution Alignment for Domain Adaptation with Pre-trained Inception ResNet.

If you have any questions, please email to yoz217@lehigh.edu

Reference

If you find it is helpful, please cite it as:

Zhang, Youshan, and Brian D. Davison. (2020). Impact of ImageNet Model Selection on Domain Adaptation. In 2020 IEEE Winter Applications of Computer Vision Workshops (WACVW). Zhang, Y., & Davison, B. D. (2019). Modified distribution alignment for domain adaptation with pre-trained inception resnet. arXiv preprint arXiv:1904.02322.

Or in bibtex style:

@article{zhang2020impact,
  title={Impact of ImageNet Model Selection on Domain Adaptation},
  author={Zhang, Youshan and Davison, Brian D},
  booktitle={In 2020 IEEE Winter Applications of Computer Vision Workshops (WACVW)},
  year={2020}
}
@article{zhang2019modified,
  title={Modified distribution alignment for domain adaptation with pre-trained inception resnet},
  author={Zhang, Youshan and Davison, Brian D},
  journal={arXiv preprint arXiv:1904.02322},
  year={2019}
}

To run the code

Open Matlab (Matlab2017a version or later should be fine)

Run ImageNet_Model_Selection.m

The newest mean accuracy of the Office+Caltech-10 dataset is around 97.5%.

Figures

ImageNet Accuracy & Parameters & Memory

T-SNE view of extracted features of Amazon domain in the Office31 dataset

T-SNE loss of sixteen neural networks of domain Amazon in the office31 dataset

Results

Correlation and R square value of Office+Caltech-10 dataset

Correlation and R square value of Office31 dataset

Correlation and R square value of Office-Home dataset