Multi-Task Learning based on Separable Formulation of Depth Estimation and its Uncertainty

Akari Asai, Daiki Ikami, Kiyoharu Aizawa

Abstract: We present an optimization framework for uncertainty estimation in a regression problem. To obtain predictive uncertainty inherent in the observation, we formulate regression with uncertainty estimation as a multi-task learning problem and a new uncertainty loss function, inspired by variational representations of robust estimation. Contrary to existing approaches, our approach allows balancing between the predictive task loss and uncertainty estimation loss. We evaluate the efficacy of our approach on NYU Depth Dataset V2 and show that our proposed method consistently yields better performance than the previous approaches, for both depth and uncertainty estimation.