Robust Semi-Supervised Monocular Depth Estimation with Reprojected\n Distances

摘要

Dense depth estimation from a single image is a key problem in computer\nvision, with exciting applications in a multitude of robotic tasks. Initially\nviewed as a direct regression problem, requiring annotated labels as\nsupervision at training time, in the past few years a substantial amount of\nwork has been done in self-supervised depth training based on strong geometric\ncues, both from stereo cameras and more recently from monocular video\nsequences. In this paper we investigate how these two approaches (supervised &\nself-supervised) can be effectively combined, so that a depth model can learn\nto encode true scale from sparse supervision while achieving high fidelity\nlocal accuracy by leveraging geometric cues. To this end, we propose a novel\nsupervised loss term that complements the widely used photometric loss, and\nshow how it can be used to train robust semi-supervised monocular depth\nestimation models. Furthermore, we evaluate how much supervision is actually\nnecessary to train accurate scale-aware monocular depth models, showing that\nwith our proposed framework, very sparse LiDAR information, with as few as 4\nbeams (less than 100 valid depth values per image), is enough to achieve\nresults competitive with the current state-of-the-art.\n