Many prediction tasks contain uncertainty. In the case of next-frame or future prediction the uncertainty is inherent in the task itself, as it is impossible to foretell what exactly is going to happen in the future. Another source of uncertainty or ambiguity is the way data is labeled. Sometimes not all objects of interest are annotated in a given image or the annotation is ambiguous, e.g. in the form of occluded joints in human pose estimation. We present a method that is able to handle these problems by predicting not a single output but multiple hypotheses. More precisely, we propose a framework for re-formulating existing single prediction models as multiple hypothesis prediction (MHP) problems as well as a meta loss and an optimization procedure to train the resulting MHP model. We consider three entirely different applications, i.e. future prediction, image classification and human pose estimation, and demonstrate how existing single hypothesis predictors (SHPs) can be turned into MHPs. The performed experiments show that the resulting MHP outperforms the existing SHP and yields additional insights regarding the variation and ambiguity of the predictions.
View on arXiv