Research On Few-shot Semantic Segmentation And Multi-object Tracking

With the popularization of smart devices and the development of Internet applications,a large amount of vision data are generated and disseminated every day,and the analysis of these data is a key issue to be solved urgently.Deep learning based methods have made major breakthroughs,and bring many conveniences to people's lives in the fields of face payment,text recognition,and unmanned driving.But deep learning based computer vision methods have several limitations: a large amount of labelled samples are required for training? the trained model cannot be directly applied to new categories.These factors have restricted the research and application of computer vision.This article is dedicated to using few-shot learning technology to solve the problem of insufficient annotation data in deep learning based computer vision methods.The research in this paper is mainly aimed at few-shot semantic segmentation and multi-object tracking.The main contributions are summarized as follows:(1)In few-shot semantic segmentation,the meta-learning based methods have received widespread attention from researchers.These works train the network by sampling support set and query set in each episode to leran the capability of class-unrelated segmentation.But these methods do not utilize the relationship between the support feature and the query feature.This paper proposes a feature contrasting based few-shot segmentation network,which utilizes the similarity between the support representation and the query representation to assist in feature extraction.(2)In multi-object tracking,the existing methods mostly adopt the design of ”trackingby-detection”,and these tracking algorithms can be divided into three modules sequentially:object detection,state estimation and data association.In these methods,the stage of object detection only contains the amount of information at the picture-level,which cannot match the video-level problem of multi-object tracking.For this reason,this paper utilizes the multiple frames information in the stage of data association,and designs a feedback-update algorithm to feedback the result of the stage of data association to the stage of detection,then updates results of the detection stage.The method proposed in this paper enhances the overall tracking ability by improving the detection accuracy.(3)We define the few-shot multi-object tracking problem firstly.This problem requires the algorithm to be able to track the category under the guidance of a few labelled data at unseen categories.In order to solve this problem,this paper proposes a transfer-learning based fewshot multi-object tracking method.This method uses basic categories of dataset to train the basic tracking model,and then utilizes a few number of labelled samples of the unseen category to migrate the model ability,and finally obtains a tracking model with new category tracking capabilities.(4)In order to evaluate the multiple algorithms proposed in this paper,we conduct experiments on the several datasets.And we build a special dataset,FS-MOT,for training and evaluation on few-shot multi-object tracking.Experimental results show that the algorithms proposed in this paper have achieved significant improvements in the corresponding tasks.