This directory contains the code for paper Visual Domain Adaptation with Manifold Embedded Distribution Alignment published at ACM Multimedia conference (ACM MM) 2018 as an Oral presentation. This paper is also lucky to be ranked as Top 10 papers.

The original code is written using Matlab R2017a. I think all versions after 2015 can run the code.

For Python users, I add a MEDA.py implementation. The Python version will need to import GFK module (can be found here). However, this Python version is only for reference since the graph Laplacian (as exactly in Matlab) is not implemented.

I offer a basic demo to run on the Office+Caltech10 datasets. Download the datasets here and put the data (mat files) into the data folder.

Run demo_office_caltech_surf.m.

MEDA achieved state-of-the-art performances compared to a lot of traditional and deep methods as of 2018. The testing datasets are most popular domain adaptation and transfer learning datasets: Office+Caltech10, Office-31, USPS+MNIST, ImageNet+VOC2007.

The following results are from the original paper and its supplementary file.

Using ResNet-50 features (compare with the latest deep methods with ResNet-50 as backbone). It seems MEDA is the only traditional method that can challenge these heavy deep adversarial methods.

Download Office-31 ResNet-50 features

Using ResNet-50 features (compare with the latest deep methods with ResNet-50 as backbone). Again, it seems that MEDA achieves the best performance.

Download Office-Home ResNet-50 pretrained features

using ResNet-50 features (compare with the latest deep methods with ResNet-50 as backbone). Again, it seems that MEDA achieves the best performance.

Download Image-CLEF ResNet-50 pretrained features

• Office-31 dataset using DECAF features (compare with deep methods with AlexNet):

• Office+Caltech 10 datasets and MNIST+USPS and ImageNet+VOC:

If you use this code, please cite it as:

Jindong Wang, Wenjie Feng, Yiqiang Chen, Han Yu, Meiyu Huang, Philip S. Yu. Visual Domain Adaptation with Manifold Embedded Distribution Alignment. ACM Multimedia conference 2018.

Or in bibtex style:

@inproceedings{wang2018visual,
    title={Visual Domain Adaptation with Manifold Embedded Distribution Alignment},
    author={Wang, Jindong and Feng, Wenjie and Chen, Yiqiang and Yu, Han and Huang, Meiyu and Yu, Philip S},
    booktitle={ACM Multimedia Conference (ACM MM)},
    year={2018}
}