Unsupervised Deep Hashing with Similarity-Adaptive and Discrete Optimization

Created by Fumin Shen, Yan Xu, Li Liu, Yang Yang, Zi Huang, Heng Tao Shen

The details can be found in the TPAMI 2018 paper (release soon).

• Prerequisites
• Installation
• Usage
• Resources
• Citation

1. Requirements for Caffe, pycaffe and matcaffe (see: Caffe installation instructions).

2. Prerequisites for datasets.

   Note: In our experiments, we horizontally flip training images manually for data augmentation. If the size of your training data is small (< 100K, like CIFAR-10. MNIST), you should do this step.

   We also provide our flipping code in cifar10/flip_img.m, you can run it to handle your own datasets.

3. VGG-16 pre-trained model on ILSVC12 datasets, and save it in caffemodels directory.

Enter caffe directory and download the source codes.

    cd caffe/

Modify Makefile.config and build Caffe with following commands:

    make all -j8
    make pycaffe
    make matcaffe

We only supply the code to train 16-bit SADH on CIFAR-10 dataset.

We integrate train step and test step in a bash file train.sh, please run it as follows:

    ./train.sh [ROOT_FOLDER] [GPU_ID]
    # ROOT_FOLDER is the root folder of image datasets, e.g. ./cifar10/
    # GPU_ID is the GPU you want to train on

We supply CIFAR-10 dataset, which has been flipped. You can download it by following links:

• CIFAR-10 dataset (png format): BaiduYun.

If you find our approach useful in your research, please consider citing:

@article{'shen2018tpami',
    author   = {Fumin Shen and Yan Xu and Li Liu and Yang Yang and Zi Huang and Heng Tao Shen},
    journal  = {IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)}, 
    title    = {Unsupervised Deep Hashing with Similarity-Adaptive and Discrete Optimization},
    year     = {2018}
}