Monocular 3D Object Detection of Moving Objects Using Random Sampling and Deep Layer Aggregation

Abstract

Restoring 3D information from a single 2D image is an ill-posed problem, and accurate depth estimation is challenging. We propose a novel method that estimates depth more accurately by randomly sampling object-wise several additional points, otherwise, conventional methods mostly estimate the depth with only one center point. To sample object-wise multiple reference points, we generate sampling space, and we obtain the ground truth of several sampling points by moving the sampled points from the sampling space to the object space. Using the obtained ground truth, the depth of reference points can be obtained through the proposed reference points depth branch. The proposed network achieved better performance by allowing depth to be estimated densely rather than the conventional network estimating the depth sparsely. © 2023 IEEE.