Unsupervised R&R Unsupervised Point Cloud REgistration via Differentiable Rendering

differentiable alignment and rendering

• RGBD images -> feature point cloud -> correspondence -> alignment -> rendering
  • feature: coord, color, feature
  • correspondence from cosine distance in features
  • ratio test to drop some estimations
  • alignment by Kabsch’s algorithm
  • randomized optimizations by random start point
• forcing feature alignment: render this view by another view’s feature point cloud