Shujie Chen, Zhonglin Liu, Jianfeng Dong, Xun Wang, Di Zhou
{"title":"Representation alignment contrastive regularisation for multi-object tracking","authors":"Shujie Chen, Zhonglin Liu, Jianfeng Dong, Xun Wang, Di Zhou","doi":"10.1049/cvi2.12331","DOIUrl":null,"url":null,"abstract":"<p>Achieving high-performance in multi-object tracking algorithms heavily relies on modelling spatial-temporal relationships during the data association stage. Mainstream approaches encompass rule-based and deep learning-based methods for spatial-temporal relationship modelling. While the former relies on physical motion laws, offering wider applicability but yielding suboptimal results for complex object movements, the latter, though achieving high-performance, lacks interpretability and involves complex module designs. This work aims to simplify deep learning-based spatial-temporal relationship models and introduce interpretability into features for data association. Specifically, a lightweight single-layer transformer encoder is utilised to model spatial-temporal relationships. To make features more interpretative, two contrastive regularisation losses based on representation alignment are proposed, derived from spatial-temporal consistency rules. By applying weighted summation to affinity matrices, the aligned features can seamlessly integrate into the data association stage of the original tracking workflow. Experimental results showcase that our model enhances the majority of existing tracking networks' performance without excessive complexity, with minimal increase in training overhead and nearly negligible computational and storage costs.</p>","PeriodicalId":56304,"journal":{"name":"IET Computer Vision","volume":"19 1","pages":""},"PeriodicalIF":1.5000,"publicationDate":"2024-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1049/cvi2.12331","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"IET Computer Vision","FirstCategoryId":"94","ListUrlMain":"https://onlinelibrary.wiley.com/doi/10.1049/cvi2.12331","RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q4","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0
Abstract
Achieving high-performance in multi-object tracking algorithms heavily relies on modelling spatial-temporal relationships during the data association stage. Mainstream approaches encompass rule-based and deep learning-based methods for spatial-temporal relationship modelling. While the former relies on physical motion laws, offering wider applicability but yielding suboptimal results for complex object movements, the latter, though achieving high-performance, lacks interpretability and involves complex module designs. This work aims to simplify deep learning-based spatial-temporal relationship models and introduce interpretability into features for data association. Specifically, a lightweight single-layer transformer encoder is utilised to model spatial-temporal relationships. To make features more interpretative, two contrastive regularisation losses based on representation alignment are proposed, derived from spatial-temporal consistency rules. By applying weighted summation to affinity matrices, the aligned features can seamlessly integrate into the data association stage of the original tracking workflow. Experimental results showcase that our model enhances the majority of existing tracking networks' performance without excessive complexity, with minimal increase in training overhead and nearly negligible computational and storage costs.
期刊介绍:
IET Computer Vision seeks original research papers in a wide range of areas of computer vision. The vision of the journal is to publish the highest quality research work that is relevant and topical to the field, but not forgetting those works that aim to introduce new horizons and set the agenda for future avenues of research in computer vision.
IET Computer Vision welcomes submissions on the following topics:
Biologically and perceptually motivated approaches to low level vision (feature detection, etc.);
Perceptual grouping and organisation
Representation, analysis and matching of 2D and 3D shape
Shape-from-X
Object recognition
Image understanding
Learning with visual inputs
Motion analysis and object tracking
Multiview scene analysis
Cognitive approaches in low, mid and high level vision
Control in visual systems
Colour, reflectance and light
Statistical and probabilistic models
Face and gesture
Surveillance
Biometrics and security
Robotics
Vehicle guidance
Automatic model aquisition
Medical image analysis and understanding
Aerial scene analysis and remote sensing
Deep learning models in computer vision
Both methodological and applications orientated papers are welcome.
Manuscripts submitted are expected to include a detailed and analytical review of the literature and state-of-the-art exposition of the original proposed research and its methodology, its thorough experimental evaluation, and last but not least, comparative evaluation against relevant and state-of-the-art methods. Submissions not abiding by these minimum requirements may be returned to authors without being sent to review.
Special Issues Current Call for Papers:
Computer Vision for Smart Cameras and Camera Networks - https://digital-library.theiet.org/files/IET_CVI_SC.pdf
Computer Vision for the Creative Industries - https://digital-library.theiet.org/files/IET_CVI_CVCI.pdf