A new Evaluation Approach for Deep Learning-based Monocular Depth Estimation Methods

Abstract

In smart mobility based road navigation, object detection, depth estimation and tracking are very important tasks for improvement of the environment perception quality. In the recent years, a surge of deep-learning based depth estimation methods for monocular cameras has lead to significant progress in this field. In this paper, we propose an evaluation of state-of-the-art depth estimation algorithms based on single single input on both the KITTI dataset and the recently published NUScenes dataset. The models evaluated in this paper include an unsupervised method (Monodepth2) and a supervised method (BTS). Our work lies in the elaboration of novel depth evaluation protocols, object depth evaluation and distance ranges evaluation. We validated our new protocols on both KITTI and NUScenes datasets, allowing us to get a more comprehensive evaluation for depth estimation, especially for applications in scene understanding for both road and rail environment.

A Comparative Study of Deep Learning-based Depth Estimation Approaches: Application to Smart Mobility

“…The objective of this study is to offer qualitative and quantitative evaluation of the state-of-the-art methods for road environments. This work is in line with the work we have carried out on the perception of the environment for the autonomous vehicle [5], [6]. The remainder of this paper is organized as follows.…”