An automotive case study on the limits of approximation for object detection

The accuracy of camera-based object detection (CBOD) built upon deep learning is often evaluated against the real objects in frames only. However, such simplistic evaluation ignores fact that many unimportant are small, distant, or background, and hence, their misdetections have less impact than those for closer, larger, foreground domains as autonomous driving. Moreover, sporadic irrelevant since confidence on detections typically averaged across consecutive frames, devices (e.g. cameras, LiDARs) redundant, thus providing fault tolerance. This paper exploits intrinsic tolerance CBOD process, assesses an automotive case study to what extent can tolerate approximation coming from multiple sources lower precision arithmetic, approximate arithmetic units, even random faults due to, instance, low voltage operation. We show within 1% baseline when considering three simultaneously, be exploited build highly efficient accelerators cars.