SD-Depth: Light-Weight Monocular Depth Estimation Using Space Depth CNN for Real-Time Applications

Abstract

With the help of the space-to-depth and depth-to-space modules, we provide a convolutional neural network design for depth estimation. We show designs that down sample the spatial information of the picture utilizing space-to-depth (SD) as opposed to the widely used pooling methods (Max-pooling and Average-pooling). The space-to-depth module may shrink the image while maintaining the spatial information of the image in the form of additional depth information. This technique is far superior to Max-pooling, which diminishes the image’s information and features. We also suggest a lightweight decoder step that builds a high-resolution depth map out of many low-resolution feature maps using the depth-to-space (DS) module. The suggested architecture effectively learns depth estimation with high processing speed and accuracy. We trained and evaluated our suggested model on NYU-depthV2 dataset and attained low error values (RMSE=0.342) and high delta accuracies (δ3=0.996) at a fast-processing speed (25Fps).