Multi-view 3D Models from Single Images with a Convolutional Network

model

• input: image and desired viewpoint
  • viewpoint: azimuth, elevation(cos and sin), distance from object center
• output: image under desired viewpoint and depth image
• architecture: simple encoder-decoder architecture
• dataset: random viewpoint and location rendered shapenet models on random background image
  • panda3d renderer?

what’s learned in latent variables?

• t-SNE: shape and color
• interpolating latent variables