Single-view to Multi-view Reconstructing Unseen Views with a Convolutional Network

encoder-decoder network

• trained on randered Shapenet models
  • generlizes well on natual images
    • well…semantically labelled natural images
• predict depthmap besides color images
  • 3D information… visualized by t-SNE on hidden embeddings

lit. review

• image transformation
  • DBM, gated autoencoder: do not scale to large images
  • transforming autoencoder: do predefined transformation
  • multi-view perceptron:
  • interpolate between similar models
• 3D model from single image
  • keypoint annotation
  • images from simalar models
  • need explicit 3D model

model

• input image ==> CCCCCL, desired vp ==> LLL
• concat, LLLCCCCC

drawback

• fail to capture finer details
  • high varience in distribution of details in dataset

another importance of best view selection?

• provide the most informative view for novel view synthesis
• and for 3D reconstruction