Adversarial Multiple Hypothesis Prediction

In a recent work we studied the possibilities of predicting multiple hypotheses with a single CNN. This helps in cases where the outcome is not certain and multiple outcomes are possible. (see our arxiv paper for details). One of the applications that we were investigating was video prediction, that is predicting a future frame of a video sequence. Recently, adversarial learning has been shown to be able to predict high quality images by learning the loss instead of explicitly defining it. We believe that a combination of adversarial ideas with our MHP model could result in better images and multiple different predictions for the future.

Supervision: Christian Rupprecht, Iro Laina, Federico Tombari


For further information please contact: Christian Rupprecht