Compact Depth-wise Separable Precise Network For Depth Completion

Abstract

Predicting a dense depth map from LiDAR scans and synced RGB images with a small deep neural network is a challenging task. Most top-accuracy methods boost precision by having a very large number of parameters and as a result huge memory consumption. Whereas, depth completion tasks are commonly tackled in areas like autonomous driving such that edge devices, which are powered by embedded GPU, are the main platform. In this work, we present a methodology to produce an efficient depth completion model with high-performance fidelity inspired by PENet. To create a compact model, we propose replacing convolutional encoder layers with depth-wise separable convolution and transpose convolutional decoder with up-sampling plus depth-wise separable convolution. Our technique of using random layer pruning as a stability test guilds the design of the architecture and avoids over-parameterization. We also provide a simple but robust knowledge distillation method to further turbocharge the network and make our model scalable to fulfill a better quality requirement. Our experiments have shown a decent improvement over the leading compact models while using significantly fewer parameters compared with other larger models.