Joint Grouping And Labeling Based On Deep Subgraph Decomposition

Joint grouping and labeling of multi-object is an extremely important basic task in computer vision and pattern recognition which has high research value and wide application in medical monitoring,man-machine interaction,intelligent transportation,content-based video retrieval and so on.Research on single tasks such as multi-object grouping and target classification has made progress,but there is limited research on joint grouping and labeling.In recently years,scholars have put forward many solutions on single tasks,but overall,the results are not satisfying on joint tasks.Currently,most research on action recognition is based on single action recognition.There is fewer research based on multiple person action recognition than those on single person.In the research of recognition single person action,the relation between interaction is ignored,and those method recognized as a general action is usually unable to express intrinsic properties of the interaction,hence the accuracy is not satisfying.Usually the external factors such as camera motion,illumination and complex of inner action will have a side effect on the human interaction recognition,which increase the difficulty of building model,and limit the recognition and understanding of multiple people actions.In this paper,we propose the task of joint grouping and labeling based on deep subgraph decomposition without knowing the number of grouping beforehand,which is different from other clustering algorithms.The method proposed in the latter section connects different target with complete graph,and decomposes them into complete subgraph which represents different groups.As a result of this,the grouping of action using this method can improve the accuracy of the human recognition and make it widely applicable.Joint grouping and labeling of multi-object is the research hotspot which has wide application.Despite the progress has been made on this task,there are some difficulties such as the building of the model architecture,choosing of the effective features,and analyzing of subgraph decomposition theory.We propose a novel method applying the framework of mixing deep feature,rich contextual clue and parameters learned from data,and uses complete graph as background graph to initialize cost function.We then propose alternating search algorithm to solve the relevant inference problem efficiently.Not only can it solve problem of human interaction,but also it can deal with the grouping and labeling on fashion group based on dominant color.Experiments were carried on three human interaction recognition datasets and a fashion group dataset to evaluate the effectiveness of our method.The result shows the method proposed in this paper is able to group and label accurately,and demonstrates our method outperforms the state-of-the-art on these datasets.