Analyses of Time Efficiency and Speed-ups in Inference Process of Two-Stage Object Detection Algorithms
Jiang Wu, Yifeng Sun, Guangming Tang, Xiaoyu Xu · 2018
Due to the good performance of the two-stage object detection algorithms in accuracy, based on a typical two-stage method-Faster RCNN, this paper combs its framework, analyzes the time consumption and proposes four speed-ups-reducing the network layers, reducing the neurons, inputting multi-scale image and setting score threshold before sorting to reduce the computational complexity. Using the same PASCAL VOC 2007 datasets, the speed-ups proposed in this paper can improve the detection speed by 35% while keeping high detection accuracy.