Location Confidence Learning In Object Detection And Segmentation

Object detection and segmentation is the basic research task of computer vision,and also the basis of many important applications,such as automatic driving,face recognition,security monitoring.In object detection and segmentation,each result of the trained model output includes object category,location,and a confidence(probability).This confidence is an important basis for judging the reliability of the results of object detection and segmentation in practical application.Accurate estimation of the confidence of object detection and segmentation is helpful to improve the robustness of the relevant application systems.However,there are still many problems in the current confidence estimation in object detection and segmentation.For example,the mismatch between the confidence and the location of the object,and highly overlapping objects are suppressed by some high confidence objects in the Non-Maximum Suppression(NMS)process.This paper proposes corresponding solutions to these problems.1.To solve the problem of mismatch between the confidence and the location of the object,this paper proposes a confidence learning method based on the score of candidate location,in which the candidate is the output results of the model.At present,most of the confidence learning methods only consider the category information of the object,but not the location accuracy of the object(that is,whether the location of the object is accurate or not),which leads to the problem of mismatch between the confidence and the object location.In this paper,both of them are considered at the same time,and an evaluation model of location accuracy is proposed for predicting the location accuracy,then the classification information and location accuracy of the object are fused,the object is scored again,and the original confidence is corrected.Our method can be simply added to the existing algorithm framework and can bring significant accuracy improvement.2.Aiming at the problem of high overlapping objects being suppressed in the NMS process,this paper proposes a confidence learning method based on the location similarity between candidates.In the NMS process,Intersection over Union(Io U)is regarded as the similarity between the candidates in the existing methods,which is unreasonable in the case of high object overlap.In this paper,a new similarity measurement method is proposed,each candidate is given an embedding feature,which fully considers the position relationship between objects,and learns the similarity between candidates through the embedding feature,and decide whether to suppress the confidence of the candidates,experiments show that our method can retain more accurate object candidates and improve the accuracy of object detection in the case of high object overlap.Based on the main line of location confidence learning,this paper mainly solves the two problems mentioned above.Experiments show that our proposed method can solve the above problems well and improve the accuracy obviously.