FacialLandmark Caffe Save

What's this

1. Facial Landmark Detection with Caffe CNN.
2. Implementation of the two nets from different paper.

• Net1: Approaching human level facial landmark localization by deep learning
• Net2: TCNN: Facial Landmark Detection with Tweaked Convolutional Neural Networks

How to prepare dataset

• Please download CelebA dataset
• change input path to your dataset path in train.prototxt file.

How to training

Please run these command in the root of this project. Note: Please change parameter in train.prototxt. such as process num and buffer2memory!

• set env

export PYTHONPATH=`pwd`/common/:$PYTHONPATH

• train net1

caffe train --solver=training/net1/solver.prototxt --gpu=0

• train net2

caffe train --solver=training/net2/vanilla_adam_solver.prototxt --gpu=0

How to testing or predict

cd testing
python test.py ../model/net1/_iter_100000.caffemodel ../training/net1/deploy.prototxt

Please replace correct path of caffe model and deploy file in above command.

How to do a benchmark

cd benchmark
python test.py ../model/net1/_iter_100000.caffemodel ../training/net1/deploy.prototxt

Please replace correct path of caffe model and deploy file in above command.
After do that, you will get mean error of your model.

Reference

VanillaCNN

Other DL framework

• The tensorlfow implementation to see here
• The mxnet implementation to see here