Learning to Generate Textures on 3D Meshes

approach

• inputs: normalized object coordinates, masks, albedo
  • avalible from blender for me
• data: unpaired 2d/3d data
• from 2.5D information to multiview images, then generate texture using diff. renderer
  • first stage: cyclegan-style
  • seconds stage: L_2 + perception loss + consistancy loss

comparable result to novel view synthesis

• but only on one dataset: CompCar
• handle multiclass scenario?
• comparation to geometry method?