Accelerates inference by separable max-pooling for object detection. (11th March 2023)

• Record Type: Journal Article
• Authors: Yang, Zhen; Ouyang, Zhikui; Yang, Fan; Yin, Zhijian
• Is Part Of: International journal of system control and information processing. Volume 4:Number 1(2022)
• Page Start: 43
• Page End: 55
• Publication Date: 2023-03-11
• Subjects: neutral network -- object detection -- YOLOX -- deep learning -- ACSPP; System design -- Data processing -- Periodicals; Information technology -- Periodicals; 003.5
• Languages: English
• ISSNs: 1759-9334

Abstract

The object detector can achieve better performance, but it also requires high computational cost. In this work, we present an accelerated inference in the structure spatial pyramid pooling (SPP), called ACSPP, based on an optimised architecture that uses combinational separable max-pooling to replace the standard max-pooling operator, and apply it to the YOLOX model. For the MS COCO Val dataset, the average inference time the proposed ACSPP is 10.87 ms, while the average inference time of the YOLOX model is 12.46 ms when a single NVIDIA 1660ti GPU is used. For the Pascal VOC2007 val dataset, the average inference time of the proposed ACSPP is 33.35 ms, while the average inference time of the YOLOX model is 37.30 ms when a single NVIDIA 1050ti GPU is used.