Deep Pedestrian Detection With Variable Input Size

In recent years,the object detection based on deep learning shows very good performance.With the input images,the neural networks learn the method to extract the feature maps from the original images end-to-end.However,the detection algorithms based on the deep learning are mainly dependent on the convolution neural network which requires a fixed size input because of the final full connection layer.So the previous deep learning algorithms need to crop/warp the original image to a fixed size,these operation greatly limit the performance on large and small objects.Based on the size problem in the object detection algorithm,this paper proposes the variable input scale object detection algorithm.The details are summarized as follows:(1)Based on the problem of fixed size dependency in convolution neural networks,this paper proposes region proposal framework using multiple scale inputs for training and testing.The framework works well for the large and small objects in the pictures.The proposed algorithm generates the regions based on the size in different levels.And it achieves good performance via the adaptive strategy.With our method,the performance on small and large objects has a significant advance.Based on the further research in the first problem,the problem of hard examples mining is studied.An associated work network,which contains a metric coding net(MC-net)and a weighted association CNN(WA-CNN),is introduced to mine the hard negatives.MC-net is based on the metric learning theory for the comparability determination,to strengthen the difference of intra-class.WA-CNN can be regarded as a network to reinforce the distance of inter-class and associates the MC-net to accomplish the detection task by a weighted strategy.Finally,because the pedestrian detection needs to process bounding boxes to carry out the final target box,Non-maximum Suppression(NMS)is used usually.But in reality,pedestrians have a high overlap in the picture,and the NMS is not work well in these scenes.This paper proposes the segment non-maximum suppression strategy for these scenes.(2)Based on the analysis in the one-stage detection and two-stage detection,this paper synthesizes the advantage of the binary classification and multi-classification in the detection and designs a one-stage detection algorithm with variable input size.Future,an Non-maximum Suppression strategy jointing multi-class score with the binary classification score,is introduced in the paper.