Deep Supervised Hashing for Image Retrieval

This is non-official Torch implementation with minor modification for Deep Supervised Hashing for Fast Image Retrieval.

The official Caffe implementation can be found here.

Prerequisites

• Install Torch

• Datasets

  • Install MNIST: luarocks install mnist
  • Install Fashion-MNIST: luarocks install https://raw.github.com/mingloo/fashion-mnist/master/rocks/fashion-mnist-scm-1.rockspec
  • CIFAR10: the code could directly download this dataset.

• matio: we use matlab scripts that adopted from here to calculate retrieval Mean Average Precision, mAP.

• weight-init.lua is adopted from here.

Usage

Please run the run.sh for training model from scratch and obtaining final hashing representation with default parameter settings for CIFAR10/MNIST/Fashion-MNIST datasets:

bash run.sh

Then copy the output result with suffix .mat in output/ directory to the scripts/ directory and run DSH.m to calculate the retrieval mAP.

Results