Relighting using Neural Texture and GAN

This project use GAN and Neural Texture to synthesis the image on new view and light condition.

Pipeline

Training data: the object under 32 view directions and 384 lighting directions, a total of 12288 photos.

Results

Novel View Point Synthesis:

Relighting:

Usage

1. Copy the data to folder data/
   • Check stereo/ directory and file mask_cam00.png
   • Copy normal map, lights_8x8.bin to data/
2. Copy the mesh to D:\Code\Project\temp\PathTracer\PathTracer
3. Run gen_uv_map.py to generate uv_map
4. Write a config file:
   • Modify OBJECT_NAME
   • Modify DATASET.NAME
   • Modify DATASET.ROOT
5. Run:

python train_gan_new.py --cfg <yaml config file>