Ulreich, Fabian (Technische Hochschule Ingolstadt), Kaup, Andre (University of Erlangen-Nuremberg), Ebert, Martin (Technische Hochschule Ingolstadt)

Novel Test Bench for End-To-End Validation of Monocular Depth Estimation under the Influence of Glaring Situations

Scheduled for presentation during the Regular Session "Sensing, Vision, and Perception IV" (ThBT5), Thursday, September 26, 2024, 15:50−16:10, Salon 13

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

The black-box nature of deep learning models employed in automated driving functions requires suitable evaluation tools. Efforts are being made to increase the validity of testing environments for real-world operations. Understanding the impact of the sensor characteristics and degradation on the downstream task of perception is another field of research. We propose a test environment for vision-based autonomous driving functions in which a real camera and a deep learning model can be evaluated jointly. Our approach enables the validation under real-world brightness conditions through projector technology. To demonstrate its applicability, we employed a Vision Transformer to perform monocular depth estimation. Our experimental setup included a challenging scenario involving glare to assess the differences in performance between the testing environments: camera test bench and simulation. We quantified the gap by contrasting image quality metrics of partly-synthetic and pure synthetic data with real-world data contained in the KITTI depth dataset. With our approach, we were able to produce images that are 37% closer to real-world than synthetic image data. Also, the gap in data variability is 18% less than with synthetic data. In addition, we found that clipping in glare situations does not necessarily lead to large errors in depth prediction.