Despite the presence of many different benchmark datasets and metrics, there is a lack of consistent perception requirements. This work focuses on the task of 3D object detection which aims at perceiving location and attributes of dynamic objects in space. Most notably, there is a lack of consistent definitions of pass and fail criteria for any given perception metric. This work addresses this topic by systematically considering human performance across a variety of visual tasks. This approach yields interpretable perception metrics as well as thresholds for pass/fail criteria. A validation approach leveraging a prediction network is introduced and successfully applied to the criteria. Comparisons with existing detectors show that current perception algorithms exhibit failures for a majority of objects on the nuScenes dataset. This indicates the necessity of explicit safety consideration in the development of perception algorithms for the automated driving task.