Decoupling Classification And Localization In Object Detection

Object detection is an important research field in computer vision,which plays an indispensable role in face detection,vehicle detection and many other applications.With the development of deep learning technology,compared with traditional target detection methods,object detection algorithm,deep learning-based object detection methods have made great progress in algorithm accuracy.Compared with image recognition task,object detection not only needs to recognize the categories of the existing targets in the image,but also needs to regress and predict the location of the object in the image.However,in the current mainstream object detection algorithms,such as Faster RCNN,they still suffer from the phenomenon of feature coupling.Specifically,the classification and regression parts within the network share the same parameters,which makes the features of these two parts are highly coupled.In view of the above problems,this thesis first analyzes the influence of the coupling of classification and regression features in the object detection algorithm,Faster RCNN as an example.The specific consequences are as follows:(1)the classification score are not consistent with the regression prediction,that is,the bounding box with highest classification score do not necessarily output the most accurate bounding box.(2)The features suitable for classification are not necessarily suitable for regression location;similarly,the features suitable for location are not necessarily the most discriminate features of the object.To solve these two problems,two novel head networks are designed to decouple the classification and regression features.Firstly,aiming at the inconsistency between classification features and regression prediction,this thesis designs a classification regression consistency prediction network based on Faster RCNN.Specifically,in the process of classification and regression,this thesis first regresses the bounding box once,then extracts the features from the predicted bounding box after regression.Hence,the features used in classification are consistent with the prediction of regression,which effectively alleviates the problem of inconsistency between classification and bounding box prediction.Secondly,aiming at the problem of feature coupling in the classification and regression part of Faster RCNN,this thesis designs a task decoupled feature extraction method.Specifically,this thesis divides the classification and regression into two sub networks.For the regression part,we use the coordinate convolution method to introduce more abundant spatial knowledge.In the classification part,deformable convolution is used to involve better adaptability to the object with different scales and shapes.Finally,based on the above two methods,experiments are carried out to verify the effectiveness of the algorithm on MS-COCO,which is the mainstream data set in the field of object detection.Compared with the Faster RCNN baseline network,the proposed method brings about 3.2% improvement in the m AP,and also shows advantages compared with other similar methods.In addition,ablation experiments are carried out to verify the effectiveness of each component.