Vehicle Detection Based on Separable Reverse Connected Network
Vehicle detection is a challenging problem which plays an important role in a wide range of traffic applications.In this paper,we propose a fast and accurate framework for detection vehicles on multi-scale using Separable Reverse Connected network.Reverse connected structure enriches the semantic information of former layers,while separable convolution is introduced for reducing computation costs.Further,optimization methods based on multi-scale training and model compressing are employed to make training process more efficient and reduce the parameters of the network with slightly loss of accuracy.Comprehensive evaluations on Pascal VOC 2007+2012 and MSCOCO 2014 show that proposed method outperforms that using Feature Pyramid Network(FPN)about 3%in mAP of vehicle categories.Model compressing accelerates the network of two-stage detector about 10 times without distinct drop of accuracy,which brings about high quality for real-time vehicle detection.
Vehicle detection Separable Reverse Connected network Model compression Convolutional neural networks
Enze Yang Linlin Huang Jian Hu
Beijingjiaotong University,Beijing,China
国际会议
广州
英文
138-149
2018-11-23（万方平台首次上网日期，不代表论文的发表时间）