Edge-preserving Near-light Photometric Stereo with Neural Surfaces

Dependencies

The proposed method is implemented in PyTorch.

• Python 3.7
• PyTorch (version = 1.9.1)
• numpy
• scipy
• CUDA-9.0
• Pyvista
• Matplotlib
• opencv-python

Overview

We provide:

• Datasets and results:
  • Real-captured near-light image data
  • Synthetic near-light image data with ground truth surface normal and depth
• Estimation results from existing methods and ours shown in the main paper
• Implementing of our method
  • modules.py: Network structure of our neural surface
  • loss_functions.py: Reconstruction loss with albedo depending on the surface normal and depth
• Code to reproduce the experimental results shown in the paper
  • demo.py

Get started

• Download the data and result into the data folder and unzip

• Check the data and released results from existing methods and ours, e.g.

  • synthetic_data
    • Buddha
      • render_img: record rendered image data
      • render_para: record GT surface normal, depth, 3D mesh, point light positions and radiant parameters
      • Released_Result: save recovered surface normal and depth, reconstructed 3D mesh

• Reproduce the experimental results shown in the paper

  python demo.py
  

  The shape estimation results from our method will be saved at

  • ./data/synthetic_data/objectname/Result_Ours/datetime_TPAMI_submit_experimentname/Recoverd_Shapes
  • ./data/real_data/objectname/Result_Ours/datetime_TPAMI_submit_experimentname/Recoverd_Shapes

The network structure follows SIREN network.