CDVAE: Co-embedding Deep Variational Auto Encoder for Conditional Variational Generation. (arXiv:1612.00132v1 [cs.CV])

Problems such as predicting an optical flow field (Y) for an image (X) are ambiguous: many very distinct solutions are good. Representing this ambiguity requires building a conditional model P(Y|X) of the prediction, conditioned on the image. It is hard because training data usually does not contain many different flow fields for the same image. As a result, we need different images to share data to produce good models. We demonstrate an improved method for building conditional models, the Co-Embedding Deep Variational Auto Encoder. Our CDVAE exploits multiple encoding and decoding layers for both X and Y. These are tied during training to produce a model of the joint distribution P(X, Y), which provides the necessary smoothing. Our tying procedure is designed to yield a conditional model easy at test time. We demonstrate our model on three example tasks using real data: image saturation adjustment, image relighting, and motion prediction. We describe quantitative evaluation metrics to evaluate ambiguous generation results. Our results quantitatively and qualitatively advance the state of the art.

from cs.AI updates on arXiv.org http://ift.tt/2gMlLUj
via IFTTT