SpherePHD

Reproduce SpherePHD with python codes (https://arxiv.org/pdf/1811.08196.pdf)

Usage

• Using DataLoader.py to make data.npy and label.npy The function will sample the Sphere pixel automatically
• Get the reconstruction information from makedata.py to get division log.
• run train.py to train your own data, which includes simpleCNN and autoencoder

Implement Detail

• Sampling the pixel from panorama should be different from the original paper. I use direct projection when doing subdivsions.
• The results on Stanford2D3D and spherical MNIST is close to the original paper

Reference

example code from authors https://github.com/KAIST-vilab/SpherePHD_public