bpcpr: Cascade Pose Regression with Back Propagation

A neural network formulation for Cascaded Pose Regression (CPR) applied to Face Alignment. Implement the algorithm in [2] which is almost the same with [1], where the parameters are jointly tuned for the whole CPR algorithmic pipeline. See e.g., [3] for original CPR based face alignment.

Install

1. Install the Matlab DAG network by following the instruction therein.
2. Run the setup_path.m to add path. Note that you need modify the code before running to specify the path of the required third party toolbox.
3. Run the mex\make.m to compile the mex file. Basically it calls the nvcc compiler in the CUDA toolkit to compile the *.cu code for the mex file.

Folder Layout

tf_*.m and tfw_*.m: transformer for the Graph Transformer Network (GTN), a.k.a. DAG netowrk. In this project we view the whole network as a transformer that composites many sub-transformers

convdag_bpcpr.m: a thin wrapper of the whole network, managing training, testing, etc.

peek.m: the observer (a Design Pattern) for convdag_bpcpr.m, managing model saving, training loss plotting on the fly, etc.

mex: C and CUDA C code

util: helper functions

cache: pre-computed data

script: scripts for training

chk_rst: scripts for results inspection

References

[1]. Baoguang Shi, Xiang Bai, Wenyu Liu, Jingdong Wang, "Deep Regression for Face Alignment", arXiv:1409.5230, 2014

[2]. Peng Sun, James K. Min, Guanglei Xiong. Globally Tuned Cascade Pose Regression via Back Propagation with Application in 2D Face Pose Estimation and Heart Segmentation in 3D CT Images, http://arxiv.org/abs/1503.08843, 2015

[3]. Ren, Shaoqing, Cao, Xudong, Wei, Yichen, and Sun, Jian. Face alignment at 3000 fps via regressing local binary features. CVPR 2014