Awesome
Deep Hash Distillation for Image Retrieval (ECCV2022)
Official Pytorch implementation of "Deep Hash Distillation for Image Retrieval" Accepted to ECCV2022 - <a href="https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136740345.pdf">DHD</a>
Overall training procedure of DHD
<p align="center"><img src="figures/framework.png" width="900"></p>Requirements
Prepare requirements by following command.
pip install -r requirements.txt
Train DHD models
Prepare datasets
We use public benchmark datasets: ImageNet, NUS-WIDE, MS COCO.
Image file name and corresponding labels are provided in ./data.
Datasets can be downloaded here: <a href="https://drive.google.com/file/d/1TAjFKnOEse4xU_ScZOM8NgQLGexebmRn/view?usp=share_link">NUS-WIDE</a> / <a href="https://drive.google.com/file/d/1EsRZP3YsLbkbJ9rNXA4x5BFkHVFIGlQP/view?usp=share_link">MS COCO</a>
For ImageNet, please download through official website <a href="https://www.image-net.org/download.php">ImageNet</a> and follow our data configuration.
Example
- Train DHD model with ImageNet, AlexNet backbone, 64-bit, temperature scaling with 0.2
python main_DHD.py --dataset=imagenet --encoder=AlexNet --N_bits=64 --temp=0.2
python main_DHD.py --help will provide detailed explanation of each argument.
Retrieval Results with Different Backbone
S: Swin Transformer, R: ResNet, A: AlexNet
ImageNet
<p align="center"><img src="figures/Imagenet_results.png" width="900"></p> NUS-WIDE <p align="center"><img src="figures/Nuswide_results.png" width="900"></p> MS COCO <p align="center"><img src="figures/Mscoco_results.png" width="900"></p>Citation
@inproceedings{DHD,
title={Deep Hash Distillation for Image Retrieval},
author={Young Kyun Jang, Geonmo Gu, Byungsoo Ko, Isaac Kang, Nam Ik Cho},
booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
year={2022}
}