Abstract

Learned confidence measures gain increasing impor-tance for outlier removal and quality improvement in stereovision. However, acquiring the necessary training data istypically a tedious and time consuming task that involvesmanual interaction, active sensing devices and/or syntheticscenes. To overcome this problem, we propose a new, flexi-ble, and scalable way for generating training data that onlyrequires a set of stereo images as input. The key idea ofour approach is to use different view points for reason-ing about contradictions and consistencies between multi-ple depth maps generated with the same stereo algorithm.This enables us to generate a huge amount of training datain a fully automated manner. Among other experiments,we demonstrate the potential of our approach by boost-ing the performance of three learned confidence measureson the KITTI2012 dataset by simply training them on avast amount of automatically generated training data rather than a limited amount of laser ground truth data.