Joint Embeddings of Shapes and Images via CNN Image Purification

@article{li2015joint,
  title={Joint embeddings of shapes and images via cnn image purification},
  author={Li, Yangyan and Su, Hao and Qi, Charles Ruizhongtai and Fish, Noa and Cohen-Or, Daniel and Guibas, Leonidas J},
  journal={ACM transactions on graphics (TOG)},
  volume={34},
  number={6},
  pages={1--12},
  year={2015},
  publisher={ACM New York, NY, USA}
}

• Have identical latent space for shapes and images.
• So retrieval would just map the input into the latent space and find nearest neighbors

1. First map the shapes into a latent space
2. Second fit the images to the mapping
   • Done with CNN
   • can be done since: Known correspondences between images and shapes, since images are just the same shapes rendered with enviroment map

• Not trained End to End