Unsupervised Monocular Depth Estimation with Encoder-decoder Network

Although fully supervised learning methods could achieve good results for monocular depth estimation, it is very expensive to obtain ground truth depth data for training. In this paper, an encoder-decoder network is proposed to estimate monocular depth with unsupervised learning. During training, this network is able to enhance the weight of the effective features in the image by learning. Therefore, the encoder is able to extract more effective features and the decoder generates more targeted results. Compared to supervised learning, the proposed network is trained by stereo pairs and image reconstruction error, rather than ground truth depth. The proposed network achieves competitive results for monocular depth estimation on the KITTI driving dataset.