Semi-supervised Depth Estimation from a Single Image Based on Confidence Learning

Abstract

Recent advances in deep convolutional neural network have lead to significant improvement over depth estimation from a single image. However, training such networks usually needs a large amount of high-quality labeled training data which is difficult to collect. To alleviate it, a semi-supervised method based on confidence learning is proposed to improve the results of depth estimation with additional unlabeled data. We utilize the confidence map generated by a confidence network to predict the trusty regions of depth estimations on unlabeled data. These depth estimations are taken as pseudo ground truth and used to train the depth network together with labeled data. Experiments on NYU Depth dataset V2 show our proposed semi-supervised method outperforms fully-supervised method.