Toward data efficient multiple object tracking
➤ Gửi thông báo lỗi ⚠️ Báo cáo tài liệu vi phạmNội dung chi tiết: Toward data efficient multiple object tracking
Toward data efficient multiple object tracking
HO CHI MINH NATIONAL UNIVERSITYHCM CITY UNIVERSITY OF TECHNOLOGY FACULTY OF COMPUTER SCIENCE AND ENGINEERING ____________________«__________BACHELOR D Toward data efficient multiple object tracking DEGREE THESISTOWARD DATA-EFFICIENT MULTIPLE OBJECT TRACKINGCouncil : Computer ScienceThesis Advisor: Dr. Nguyen Due Dung Reviewer : Dr. Nguyen Hua Phung —0O0—Students:Phan Xuan Thanh I.am 1710163Tran Ho Minh Thong 1710314HO CHI MINH CITY. 07/2021DeclarationWe declare that the thesis entitled ■■ TOWA Toward data efficient multiple object tracking RD DATA-Eb'EIClENTMULTIPLE ORJECT TRACKING” is our own work under the supervision of Dr. Nguyen Due Dung.We declare that the information reported hereToward data efficient multiple object tracking
is the result of our own work, except where references an* made. The thesis has not been accepted for any degree and is not concurrently submitted toHO CHI MINH NATIONAL UNIVERSITYHCM CITY UNIVERSITY OF TECHNOLOGY FACULTY OF COMPUTER SCIENCE AND ENGINEERING ____________________«__________BACHELOR D Toward data efficient multiple object tracking as they move around a scene. MOT is an open and attractive research held with a broad extent of categories and applications such as surveillance, sports analysis, human-computer interface, biology, etc.The difficulties of this problem lie in several challenges, such as frequent occlusions, intraclas Toward data efficient multiple object tracking s and inter-class variations, etc. Recently, deep learning MOT methods confront these challenges effectively and lead to groundbreaking results. ThereToward data efficient multiple object tracking
fore, these methods are used in almost all state-of-the-art MOT algorithms. Despite their successes, deep learning MOT algorithms. like other deep leaHO CHI MINH NATIONAL UNIVERSITYHCM CITY UNIVERSITY OF TECHNOLOGY FACULTY OF COMPUTER SCIENCE AND ENGINEERING ____________________«__________BACHELOR D Toward data efficient multiple object tracking anually labeling positions of objects on every video frame (with bounding boxes or segmentation) and assigning each object to a single identity (ID), such that different objects have different IDs. and the same object in different frames has the same ID. This makes annotating MOT data a very time-co Toward data efficient multiple object tracking nsuming task.To solve the data problem in deep learning MOT algorithms, in this thesis, we will propose a method, where we only need the annotations oToward data efficient multiple object tracking
f object positions. Experiments show that our method is compatible with the current state-of-the-art method, despite the lack of object ID labeling.OnHO CHI MINH NATIONAL UNIVERSITYHCM CITY UNIVERSITY OF TECHNOLOGY FACULTY OF COMPUTER SCIENCE AND ENGINEERING ____________________«__________BACHELOR D Toward data efficient multiple object tracking integrated with MOT models, and also lack necessary features for MOT problems. Therefore, in this thesis, we will also develop a new annotation tool. Il will support automatically labeling via our proposed MOT model. Moreover, our tool will also provide plenty of convenient features, which will incr Toward data efficient multiple object tracking ease the automation for labeling processes, control the accuracy and rationality of results, and increase users’ experiences.To sum lip. our main contToward data efficient multiple object tracking
ributions in this thesis are twofold:•Our first major contribution is an MOT algorithm compatible with state-of-the-art algorithms without the need foHO CHI MINH NATIONAL UNIVERSITYHCM CITY UNIVERSITY OF TECHNOLOGY FACULTY OF COMPUTER SCIENCE AND ENGINEERING ____________________«__________BACHELOR D Toward data efficient multiple object tracking ol. Our tool supports automatic annotation and a lot of features that help fasten the labeling process of MOT data.AcknowledgmentsWe would like to thank Mr. Nguyen Due Dung for guiding US to important publications and for the stimulating questions on artificial intelligence. The meetings and convers Toward data efficient multiple object tracking ations were vital in inspiring US to think outside the box. from multiple perspectives to form a comprehensive and objective critique.Contents1IntroduToward data efficient multiple object tracking
ction11.1The Multiple Object Tracking Problem....................................... 11.2Introduction to MOT algorithms...............................HO CHI MINH NATIONAL UNIVERSITYHCM CITY UNIVERSITY OF TECHNOLOGY FACULTY OF COMPUTER SCIENCE AND ENGINEERING ____________________«__________BACHELOR D Toward data efficient multiple object tracking ........................ 51.5Thesis outline............................................................. 52Contrastive Learning & Object Detection62.1Contrastive Learning....................................................... 6 Toward data efficient multiple object tracking HO CHI MINH NATIONAL UNIVERSITYHCM CITY UNIVERSITY OF TECHNOLOGY FACULTY OF COMPUTER SCIENCE AND ENGINEERING ____________________«__________BACHELOR DGọi ngay
Chat zalo
Facebook