With the increasing development of critical systems based on artificial intelligence (AI), methods have been proposed and evaluated in academia to assess the reliability of these systems. In the context of computer vision, some approaches use the generation of images altered by common perturbations and realistic transformations to assess the robustness of systems. To better understand the strengths and limitations of these approaches, we report the results obtained on an industrial case of a road object detection system. By comparing these results with those of reference models, we identify areas for improvement regarding the robustness of the system and the metrics used for this evaluation.