Real-time Object Detection with Adaptive Image Scaling and Network Path Scheduling

Abstract

Real-time object detection is essential for autonomous vehicles. In autonomous driving, object detection should finish within a time limit for autonomous vehicles to make a right decision. Furthermore, real-time constraints can change over time depending on the dynamic execution environment. Although deep learning has remarkably enhanced detection speed and accuracy, ordinary object detectors are unaware of real-time constraints. This dissertation proposes Occamy, an object detection system that provides adaptive neural network execution for real-time object detection. Occamy exploits the trade-offs between detection time and accuracy in ordinary object detection networks to provide real-time object detection. According to the real-time constraints, Occamy dynamically changes the scale of input images and schedules the execution path of an object detection network. In evaluation, this dissertation shows that the Occamy object detection system can satisfy dynamically changing real-time constraints while minimizing accuracy loss.