Efficient Real-Time Object Detection with Adaptive Image Scaling and Cropping
TimeTuesday, December 7th6:00pm - 7:00pm PST
LocationLevel 2 - Lobby
DescriptionWe target real-time object detection for autonomous cars, aiming at providing different levels of accuracy and timeliness to image portions with different criticality levels. We propose a new computation model and DNN scheduling framework. The former decomposes a DNN into two sub-tasks: a mandatory sub-task dedicated for a safety-critical portion of an image and an optional sub-task for processing a down-scaled image. The latter enables to prioritize sub-tasks according to their criticality levels and adaptively adjust the scale of the input to meet the timing constraints while minimizing response time of mandatory sub-tasks or maximizing accuracy of optional sub-tasks.