Learning Loss for Active Learning in Depth Reconstruction Problem

Accurate depth estimation from images is a fundamental task in deep learning. It has many applications including scene understanding and reconstruction. Datasets for supervised depth estimation are hard to obtain and usually do not contain a sufficient number of images or a sufficient variety of scenes. Since inputs for depth estimation are simple RGB images, it is easy to obtain a large number of various unlabeled images. We consider that depth masks can be labeled by using manual marking. Thus, we researched the possibility of performing an active learning approach for selecting unlabeled samples to be labeled. In this work, we concentrated on using the learning loss method to perform active learning train selection. We performed multiple experiments with the learning loss algorithm and evaluated the resulting model.