dreadlord1984 / mobile-id

MobileID: Face Model Compression by Distilling Knowledge from Neurons

MobileID: Face Model Compression by Distilling Knowledge from Neurons

[Project] [Paper]

Overview

MobileID is an extremely fast face recognition system by distilling knowledge from DeepID2. Given a detected and aligned face image, this software outputs a low-dimentional face representation which faithfully preserves its identity information. More details can be found in our paper:

"Face Model Compression by Distilling Knowledge from Neurons"

Ping Luo, Zhenyao Zhu, Ziwei Liu, Xiaogang Wang, Xiaoou Tang (The Chinese University of Hong Kong)

In AAAI Conference on Artificial Intelligence (AAAI) 2016, Oral Presentation

Further information please contact Ziwei Liu.

Requirements

• Caffe
• t-SNE

Getting started

• Install and compile the Caffe and t-SNE libraries.
• Download the pre-trained model mobile_id.caffemodel:

Place "mobile_id.caffemodel" into "./models/" 

• Download the pre-stored align&&cropped LFW dataset lfw.zip:

Place "lfw.zip" into "./data/gallery/" and unzip

• Run the feature extraction script:

sh ./extract_features_gallery.sh

• Run the visualization script:

matlab ./gen_tsne_gallery.m

Performance

The MobileID system is trained on CelebA Dataset and tested on LFW Dataset. When equipped with Joint Bayesian, it achieves excellent performance as well as fast speed, as shown below:

Face verification accuracy on LFW	Runtime on CPU	Memory footprint during inference
97.32% (mean classification accuracy)	250FPS (frames per second)	2M (megabytes)

Dataset

Large-scale CelebFaces Attributes (CelebA) Dataset

Note that there are no identity overlapping between CelebA Dataset and LFW Dataset.

License and Citation

The use of this software is RESTRICTED to non-commercial research and educational purposes.

@inproceedings{luo2016mobileid,
 author = {Ping Luo, Zhenyao Zhu, Ziwei Liu, Xiaogang Wang, and Xiaoou Tang},
 title = {Face Model Compression by Distilling Knowledge from Neurons},
 booktitle = {AAAI Conference on Artificial Intelligence (AAAI)},
 month = {February},
 year = {2016} 
}