Deep Hashing for Yalefaces

This work tries to retrieve yalefaces dataset. Yalefaces is a human faces dataset which contains 165 greyscale images. We use PyTorch and the method introduced in HashNet to conduct the experiment.

Results

Method	Bits Number	mean Average Precision
HashNet	32	100%
HashNet	16	96.88%

Dataset

Original dataset contains 15 humans' faces and each category has 11 different expressions. We first crop them to 100 x 100 resolution and then randomly select 2 images in each category to make a test dataset and the train dataset is made up for the rest of them. Then we extract the HOG features of these images.

To use the data:

$ cd <Repository Root>/yuhang
$ python pre_process.py

This will generate two npz files. These two npz files also have been stored in traindata and testdata root.

HashNet

We use LeNet-5 structure with more channels to learn the hash codes. To run the training:

$ cd <Repository Root>/yuhang
$ python main.py 

Trainning 16 bits hash codes need to adjust some hyper parameter:

$ cd <Repository Root>/yuhang
$ python main.py --bitnum 16 --alpha 0.9 --lr 1e-4

The trained hash codes also have been provided. To do the evaluation:

$ cd <Repository Root>/yuhang
$ python main.py --evaluate