Amplitude Suppression and Direction Activation in Networks for 1-bit Faster R-CNN

Abstract

Recent advances in object detection have been driven by the success of deep convolutional neural networks (DCNNs). Deploying a DCNN detector on resource-limited hardware such as embedded devices and smart phones, however, remains challenging due to the massive number of parameters a typical model contains. In this paper, we propose an amplitude suppression and direction activation for Faster R-CNN (ASDA-FRCNN) framework to significantly compress DCNNs for highly efficient performance. The shared amplitude between the full-precision and the binary kernels can be significantly suppressed through a simple but effective loss, which is then incorporated into the existing Faster R-CNN detector. Furthermore, the ASDA module is generic and flexible to be incorporated into existing DCNNs for different tasks. Experiments demonstrate the superiority of 1-bit ASDA-FRCNN which achieves superior performance on various datasets. Specifically, ASDA-FRCNN shows the best speed-accuracy trade off with 63.4% at estimated 711 FPS and 19.4% mAP at and estimated 362 FPS with ResNet-18 on the PASCAL VOC 2007 and MS COCO validation datasets respectively, which demonstrate the superior performance and strong generalization of our method.