Home

Awesome

Light CNN for Deep Face Recognition

Install Required Packages

First ensure that you have installed the following required packages:

See requirements.txt for details.

Datasets

Training a Model

python train.py

You could change some arguments in the train.py, like num_epochs, gpus. If want to get the good results and save training time, there are some points should note:

Extract features of LFW

cd extract_features
python extract_features.py

The script will load the trained LightCNN model to generate fc layer output, w.t. the feature of a image. You could change the arguments in test.sh depend on your machine config. You could download the aligned LFW datasets and lfw_patch_part.txt from here download.

Face Verification

Downloading trained model

  1. Change the optimizer method, like: when loss doesn't decrease, manually decrease it to 10 time smaller.
  2. Improved by manaully aligning the images which are mis-algined in LFW datasets.
  3. Utilize a new feature of a image, for example use mfm_fc1_output as a image's feature, it will increase the accuracy in LFW.
  4. Improved by using metric learning method for similarity caculation, not just cos value.