DepthNet: A Monocular Depth Estimation Framework

Depth estimation is an important technique in computer vision for applications such as pose modelling, activity recognition, etc. Common methods of depth estimation include stereo vision which aims to trace the corresponding feature displacement to reconstruct a geometry as a depth map. To reduce the reliance on stereo vision systems, training-based deep learning methods have been utilized to generate depth maps using RGB images from a single camera. In this paper we propose a DepthNet framework to predict the relative depth of objects placed in the scene of an RGB image with respect to each other, with the help of multiple resizing upskip connections and up-convolutional layers, which further enhance its depth estimation abilities. The results specialize on the edges and the gradients of the objects existing within the scene by introduced as an Edge Loss. The single layered model with limited number of parameters which enables it to be implemented on any platform which has certain limiting factors of space occupancy and computational power. To validate the performance of the proposed architecture, the experiments are conducted on three publically available datasets: NYU depth dataset V2 [1], KITTI depth dataset [2], SUN-RGBD dataset. The projected work exhibits superior depth estimation results victimizing single RGB images, from the opposite state-of-the-arts. The performance of proposed work is evaluated on the following victimizing parameters: REL, RMS, Squared REL, and RMS_log10.