Deep Center-Based Dual-Constrained Hashing for Discriminative Face Image Retrieval (DCDH)

This repository contains the PyTorch implementation of our paper: Deep Center-Based Dual-Constrained Hashing for Discriminative Face Image Retrieval by Ming Zhang^a, Xuefei Zhe^b, Shifeng Chen^c and Hong Yan^a.

^a Department of Electrical Engineering, City University of Hong Kong

^b Tencent AI Lab

^c Shenzhen Institutes of Advanced Technology, CAS

Citation

If you find the codes are useful to your research, please consider citing our PR paper:

@article{zhang2021deep,
title = {Deep center-based dual-constrained hashing for discriminative face image retrieval},
author = {Ming Zhang and Xuefei Zhe and Shifeng Chen and Hong Yan},
journal = {Pattern Recognition},
volume = {117},
pages = {107976},
year = {2021},
}

Introduction

We propose a novel center-based framework integrating end-to-end hashing learning and class centers learning simultaneously. Unlike most existing works that are either based on pairwise/triplet labels [1, 2, 3] or softmax classification loss [4, 5], we apply a class-wise similarity as the supervision. With a normalized Gaussian-based loss function, the first constraint of DCDH minimizes the intra-class variance by clustering intra-class samples into a learnable class center. And the second constraint serves as a regularization term to enlarge the Hamming distance between pairwise class centers for inter-class separability. Furthermore, we introduce a regression matrix to encourage intra-class samples to generate the same binary codes for hashing codes compactness.

Overview

Illustration of the proposed framework. θ, M, B, and W represent the network parameters, class centers, binary hashing codes, and regression matrix, respectively. The proposed dual constraint on class centers aims to pull intra-class samples to the corresponding class center while pushing pairwise centers as far as possible. By W, B is mutually determined by the hashing layer output and labels information. The arrows between every two modules show the information forward/backward-propagation.

Prerequisites

• Python >=3.6
• PyTorch >=1.2

Data

We directly use the same data of FaceScrub and YouTube Faces datasets as released by DDH. For your convenience:

• YouTube Faces

Training: https://drive.google.com/open?id=1Of-hUKhQhk3OtCUAZc15pZqEvKHWNUpN

Test: https://drive.google.com/open?id=10RuxZuIMfvPN6ziPglJTgS6IWFBzW3pW

• FaceScrub

Training: https://drive.google.com/open?id=1STytImOvad3PgPN2NwUSVga-fOU5i2Tm

Test: https://drive.google.com/open?id=1saSeHjaJBSAUNlyC-Ismdh2VbEiyZPuU

• Subset of VGGFace2:

The processed data of the subset of VGGFace2 is also provided for your download: (total size ~2.4GB)

https://drive.google.com/open?id=1PaXIfSrw1SC1uiv4Affgm882-Q-w_icl

Models

We release the trained models on two datasets: FaceScrub and YouTube Faces under four different code lengths {12, 24, 36, 48}, respectively. The models are named following the rule of "dcdh_[dataset][length].tar" and can be downloaded here.

Usage

Training

Assuming the dataset path ./FaceScrub is under the same directory as the dcdh_train.py, to train the model by yourself, simply run the following command:

python dcdh_train.py --dataset facescrub --save your_model_name.tar --len 48

This will train the 48-bit hashing model on the dataset FaceScrub with the default configurations of hyper-parameters, which are the same as written in the paper. The model will be saved under the folder ./checkpoint in default. To see help for the training script:

python dcdh_train.py -h

Evaluation

We provide various of commonly-used evaulation metrics including mean average precision (mAP), precision and recall under two Hamming distances and the precision curve w.r.t. Hamming ranking. Suppose your model is under the default folder ./checkpoint, to evaluate, simply execute:

python evaluation.py --load model_name_on_youtube_12bit.tar --dataset youtube --len 12

Related Projects

[1] Feature Learning based Deep Supervised Hashing with Pairwise Labels (DPSH)

[2] Deep Supervised Hashing with Triplet Labels (DTSH)

[3] Deep Supervised Discrete Hashing (DSDH)

[4] Discriminative Deep Hashing for Face Image Retrieval (DDH)

[5] Discriminative Deep Attention-Aware Hashing for Face Image Retrieval (DDAH)

We also thank the above authors for releasing their codes.