Skip to content

Latest commit

 

History

History
61 lines (42 loc) · 4.36 KB

README.md

File metadata and controls

61 lines (42 loc) · 4.36 KB
Raw

Visual Explanation for Deep Metric Learning

This work explores the visual explanation for deep metric learning and its applications. We show that the proposed framework can be directly deployed to a large range of metric learning applications and provides valuable information for understanding the model. See the arxiv paper for details.

We provide runnable demos for activation decomposition and Grad-CAM (with our variants) on Image Retrieval, Person Re-identification, Face Verification, and Geo-localization. If you find this repository useful in your project, please consider citing:

@article{zhu2021visual,
  title={Visual explanation for deep metric learning},
  author={Zhu, Sijie and Yang, Taojiannan and Chen, Chen},
  journal={IEEE Transactions on Image Processing},
  year={2021},
  publisher={IEEE}
}

Requirement

- Python >= 3.5, Opencv, Numpy, Matplotlib
- PyTorch >= 1.0
- Tensorflow == 1.13.1 only for Geo-localization

Image Retrieval

Please download our pretrained model (Multi_similarity) and put it in "./Image_Retrieval/Model/". Then run the demo:

python Image_Retrieval/demo.py

You may download the CUB dataset to generate more results. Enjoy!

Person Re-identification

Please download our pretrained model (strong baseline) and put it in "./Person_Re-identification/Model/". Then run the demo:

python Person_Re-identification/demo.py

You may download the Market-1501 dataset for more results. Enjoy!

Face Verification

Please download our pretrained model (arcface) and put it in "./Face_Verification/Model/". Then run the demo:

python Face_Verification/demo.py

You may download the LFW or FIW dataset for more results. Enjoy!

Geo-localization

Please download our pretrained model package (Siamese-VGG) and put it in "./Geo-localization/Model/". Use tar -xvf model.ckpt.tar to extract. Then run the demo:

python Geo-localization/demo.py

Query pixel moving on the query street view image (left) The changing point-specific activation map on the retrieved aerial view image (right)

You may download the CVUSA dataset for more results. Enjoy! A cleaner yet stronger model is coming soon!

Reference

- https://github.com/bnu-wangxun/Deep_Metric
- https://github.com/lulujianjie/person-reid-tiny-baseline
- https://github.com/foamliu/InsightFace.git
- https://github.com/david-husx/crossview_localisation.git