Object-aware Random Cutdepth for Depth Estimation of Autonomous Driving

Abstract

Monocular Depth Estimation is a technology that predicts the distance in pixel-level by receiving an RGB single image of a camera as an input. Recently, research on depth estimation using a vision transformer have been actively conducted for monocular depth estimation, and it has surpassed performance based on a convolutional neural network. However, vision transformer has weak inductive bias compared to CNN, so it is not suitable for midium-sized dataset. To compensate for this, proper regularization and data augmentation are essential. In this paper, we proposed data augmentation suitable for depth estimation of autonomous driving dataset to compensate for the weakness. This method preserves the edge features of the RGB input image by overlaying the randomly cutdepth focusing on the object required for autonomous driving on the RGB image, and at the same time, proper regularization is performed to improve the depth estimation performance.