Pixel2Mesh++ Multi-View 3D Mesh Generation via Deformation

input: color images, camera poses

point2mesh: deform an initial shape conditioned on images

pipeline

• images -> geometry / semantic feature
• refine output of pixel2mesh
  • propose deformation for each vertex
  • assign features from image to each vertex
    • and statistics -> cross-view learning?
  • softmax of proposals -> differentiable