Fonder Michaël ; Van Droogenbroeck Marc

Abstract

When used by autonomous vehicles for trajectory planning or obstacleavoidance, depth estimation methods need to be reliable. Therefore, estimatingthe quality of the depth outputs is critical. In this paper, we show howM4Depth, a state-of-the-art depth estimation method designed for unmannedaerial vehicle (UAV) applications, can be enhanced to perform joint depth anduncertainty estimation. For that, we present a solution to convert theuncertainty estimates related to parallax generated by M4Depth into uncertaintyestimates related to depth, and show that it outperforms the standardprobabilistic approach. Our experiments on various public datasets demonstratethat our method performs consistently, even in zero-shot transfer. Besides, ourmethod offers a compelling value when compared to existing multi-view depthestimation methods as it performs similarly on a multi-view depth estimationbenchmark despite being 2.5 times faster and causal, as opposed to othermethods. The code of our method is publicly available athttps://github.com/michael-fonder/M4DepthU .