Boosting Monocular Depth Estimation with Channel Attention and Mutual Learning

摘要

We propose a novel learning-based method for monocular depth estimation with channel attention and mutual learning. First, we design a channel attention module, called cascaded channel attention module (CCAM). By applying channel attention modules in a cascade manner, CCAM produces multi-scale feature maps which are well-suited for representing 3D shapes. Then, we develop a two-branch depth prediction network (TBDP-Net) containing CCAM, and train it by mutual learning. By sharing the knowledge of each branch during training, mutual learning enables the TBDP-Net to boost the performance from the baseline which is the state-of-the-art method. Taking advantage of channel attention and mutual learning, TBDP-Net can estimate depth from feature maps which focus on features related to 3D shapes. The experimental results show that our proposed method improves the performance in all of the evaluation metrics of depth estimation with the same computational cost as the baseline.