[ICCVw19] DAME WEB: DynAmic MEan with Whitening Ensemble Binarization for Landmark Retrieval without Human Annotation
We present a compact binary global descriptor for landmark retrieval. It is 8 times smaller than the baseline GeM.
| Storage/Performance Comparison | Demo Video |
|---|---|
 |
 |
Code Author: Tsun-Yi Yang
Paper Authors: Tsun-Yi Yang, Duy-Kien Nguyen, Huub Heijnen, Vassileios Balntas
This is a demo version of the DAME WEB paper from ICCV2019 CEFRL workshop. It is done in the PhD research internship period (2019) at Scape Technologies for Tsun-Yi and Kien.
To start with running the code, you have to download the pre-trained model weight and the whitening ensemble file first Google drive
Put the folders under the main folder of this repo and it's done.
We provide ipython file and a shell script as examples for running the code (i.e. extracting global binary descriptor from a custom image).
- Shell script example
cd demo
sh run_extract.sh
- An example command for calling the python:
python3 -m TYY_extract_DAME_WEB --gpu-id '0' --network-path '../pre-trained/Res101_DAME.pth.tar' --whitening-path '../whitening/WEB_retrieval-SfM-120k.pth' --image-size 1024 --image-path '../images/big_ban.jpg' --multi_dilation False
If you use the code, please cite the following paper.
@inproceedings{yang2019dame,
title={DAME WEB: DynAmic MEan with Whitening Ensemble Binarization for Landmark Retrieval without Human Annotation},
author={Yang, Tsun-Yi and Kien Nguyen, Duy and Heijnen, Huub and Balntas, Vassileios},
booktitle={Proceedings of the IEEE International Conference on Computer Vision Workshops},
pages={0--0},
year={2019}
}
This work is heavily inspired by GeM. You may refer the training part to GeM.