Dense depth estimation based on a single image is a basic problem in computer vision and has exciting applications in many robotic tasks. Modelling fully supervised methods requires the acquisition of accurate and large ground truth data sets, which is often complex and expensive. On the other hand, self-supervised learning has emerged as a promising alternative to monocular depth estimation as it does not require ground truth depth data. In this paper, we propose a novel self-supervised joint learning framework for depth estimation using consecutive frames from monocular and stereo videos. Our architecture leverages two new ideas for improvement: (1) triplet attention and (2) funnel activation (FReLU). By adding triplet attention to the deep and pose networks, this module captures the importance of features across dimensions in a tensor without any information bottlenecks, making the optimisation learning framework more reliable. FReLU is used at the non-linear activation layer to grasp the local context adaptively in images, rather than using more complex convolutions at the convolution layer. FReLU extracts the spatial structure of objects by the pixel-wise modeling

No abstract

No abstract