A Virtual Platform for Object Detection Systems

Abstract

Computer vision is increasingly effective and important in many applications, including disease diagnosis, sports, and autonomous-driving. Visual recognition tasks, such as image classification and object detection, are the key of many of these applications, and recent developments in convolutional neural networks (CNNs) have made outstanding leaps in performance. Therefore, optimizing the data-flow between the image sensor and CNNs now constitute the majority of the effort in computer vision system design. System performance is sensitive to the qualities of the image sensor and CNN hardware accelerator. We focus on determining the influence of the sensor and accelerator on the overall performance and power of an object detection inference task. Because the relationship between image sensor quality and CNN performance is complex, we use image quality as a bridge when evaluating system performance. Developing a new product is very expensive and time consuming. This paper will offer an virtual platform for object detection systems, and each component in the system will be simulated by a proper power model and a behavior model. The power, performance, and area of the complete system will be predicted to help designers optimize object detection systems.