Surface normal prediction using stacked hourglass model

This is a course project in EECS 442 Introduction to Computer Vision at the University of Michigan, Ann Abor. A stacked hourglass model [1] is trained to predict surface normals from a single color image. All pixels where normals are to be predicted are indicated by a mask image. I reached a pixel-wise mean angle error of 0.37 and ranked 4/60 in class.

[1] A. Newell, K. Yang, and J. Deng, “Stacked Hourglass Networks for Human Pose Estimation,” arXiv:1603.06937 [cs], Mar. 2016. link