ResUnet++ for Sparse Samples-based Depth Prediction

Abstract

Depth prediction based on sparse measurements is a very effective way to obtain high-quality dense depth maps. However, most of them use a network design similar to the RGB-based approach. This results in a complex structure that does not fully exploit the sparse information. To further improve the accuracy of depth prediction, we utilize dense long and short skip connections, by expanding Unet++ to with residual block (denoted as ResUnet++), to extract more useful information from sparse depth measurements. Then, we combine deeply-supervised technique and provide a new method of lightweight network design for fast depth estimation. Experimental results verify the effectiveness of ResUnet++ on NYU-depth-v2 datasets. Combing with deeply-supervised technology, this network structure we propose can be divided into a number of subnets, which can then be used in a really flexible way to maximize the efficiency of the task.