An Information-Reserved and Deviation-Controllable Binary Neural Network for Object Detection

Object detection is a fundamental task in computer vision, which usually based on convolutional neural networks (CNNs). While it difficult to be deployed embedded devices due the huge storage and computing consumptions, binary (BNNs) can execute object with limited resources. However, extreme quantification BNN causes diversity of feature representation loss, eventually influences performance. In this paper, we propose method balancing Information Retention Deviation Control achieve effective detection, named IR-DC Net. On one hand, introduce KL-Divergence compose multiple entropy for maximizing available information. other design lightweight module generate scale factors dynamically minimizing deviation between real convolution. The experiments PASCAL VOC, COCO2014, KITTI, VisDrone datasets show that our improved accuracy comparison previous networks.