본문

서브메뉴

Continual Visual Learning for Visual Understanding
Continual Visual Learning for Visual Understanding
Continual Visual Learning for Visual Understanding

상세정보

자료유형  
 학위논문 서양
최종처리일시  
20250211153029
ISBN  
9798346857150
DDC  
621.3
저자명  
Liang, Mingfu.
서명/저자  
Continual Visual Learning for Visual Understanding
발행사항  
[Sl] : Northwestern University, 2024
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2024
형태사항  
198 p
주기사항  
Source: Dissertations Abstracts International, Volume: 86-06, Section: B.
주기사항  
Advisor: Wu, Ying.
학위논문주기  
Thesis (Ph.D.)--Northwestern University, 2024.
초록/해제  
요약Visual understanding systems have achieved remarkable success across various computer vision tasks, yet their deployment in real-world scenarios reveals a critical limitation: the inability to evolve and adapt to the dynamic nature of visual concepts. While traditional machine learning assumes a static world where all variations of visual concepts can be captured during initial training, real-world applications demand continuous adaptation to emerging visual concepts at different granularities.Despite the growing interest in continual learning, research prior to 2020 primarily focused on simple image classification tasks, leaving the unique challenges of various computer vision tasks largely unexplored. This dissertation addresses this gap by introducing a comprehensive framework for continual visual learning, defined as the systematic process of evolving a visual understanding system's capabilities across multiple granularities of visual concept evolution. This evolution encompasses three dimensions: intra-category adaptation to novel subpopulations, enhanced fine-grained semantic understanding of known objects, and incorporation of entirely new visual concepts, all while operating under real-world constraints.We present three pioneering works that advance the field of continual visual learning. First, we address the subpopulation shifting problem through incremental subpopulation learning, proposing a novel method that enables aggressive learning of new subpopulations while mitigating catastrophic forgetting of seen ones. Second, we tackle the challenge of incremental object keypoint estimation, developing a specialized framework that allows models to learn novel keypoints without access to previously trained data while improving the detection of existing keypoints. Finally, we extend beyond conventional continual learning paradigms by introducing an automatic self-improved data engine for object detection in autonomous driving, which autonomously identifies learning targets, curates relevant data, auto-labeling and continual learning with the noisy label, and validates model improvements, all powered by vision-language models and large language models.These contributions not only advance the theoretical foundations of continual learning but also address practical challenges in deploying adaptive visual understanding systems. Our work demonstrates the importance of task-specific considerations in continual learning and establishes a framework for building self-evolving computer vision systems that can autonomously adapt to the dynamic nature of real-world visual understanding tasks.
일반주제명  
Electrical engineering
일반주제명  
Computer science
키워드  
Automatic self-improve data engine
키워드  
Autonomous driving
키워드  
Continual learning
키워드  
Continual visual learning
키워드  
Incremental subpopulation learning
기타저자  
Northwestern University Electrical and Computer Engineering
기본자료저록  
Dissertations Abstracts International. 86-06B.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008250123s2024        us                              c    eng  d
■001000017164670
■00520250211153029
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798346857150
■035    ▼a(MiAaPQ)AAI31635263
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a621.3
■1001  ▼aLiang,  Mingfu.
■24510▼aContinual  Visual  Learning  for  Visual  Understanding
■260    ▼a[Sl]▼bNorthwestern  University▼c2024
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2024
■300    ▼a198  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  86-06,  Section:  B.
■500    ▼aAdvisor:  Wu,  Ying.
■5021  ▼aThesis  (Ph.D.)--Northwestern  University,  2024.
■520    ▼aVisual  understanding  systems  have  achieved  remarkable  success  across  various  computer  vision  tasks,  yet  their  deployment  in  real-world  scenarios  reveals  a  critical  limitation:  the  inability  to  evolve  and  adapt  to  the  dynamic  nature  of  visual  concepts.  While  traditional  machine  learning  assumes  a  static  world  where  all  variations  of  visual  concepts  can  be  captured  during  initial  training,  real-world  applications  demand  continuous  adaptation  to  emerging  visual  concepts  at  different  granularities.Despite  the  growing  interest  in  continual  learning,  research  prior  to  2020  primarily  focused  on  simple  image  classification  tasks,  leaving  the  unique  challenges  of  various  computer  vision  tasks  largely  unexplored.  This  dissertation  addresses  this  gap  by  introducing  a  comprehensive  framework  for  continual  visual  learning,  defined  as  the  systematic  process  of  evolving  a  visual  understanding  system's  capabilities  across  multiple  granularities  of  visual  concept  evolution.  This  evolution  encompasses  three  dimensions:  intra-category  adaptation  to  novel  subpopulations,  enhanced  fine-grained  semantic  understanding  of  known  objects,  and  incorporation  of  entirely  new  visual  concepts,  all  while  operating  under  real-world  constraints.We  present  three  pioneering  works  that  advance  the  field  of  continual  visual  learning.  First,  we  address  the  subpopulation  shifting  problem  through  incremental  subpopulation  learning,  proposing  a  novel  method  that  enables  aggressive  learning  of  new  subpopulations  while  mitigating  catastrophic  forgetting  of  seen  ones.  Second,  we  tackle  the  challenge  of  incremental  object  keypoint  estimation,  developing  a  specialized  framework  that  allows  models  to  learn  novel  keypoints  without  access  to  previously  trained  data  while  improving  the  detection  of  existing  keypoints.  Finally,  we  extend  beyond  conventional  continual  learning  paradigms  by  introducing  an  automatic  self-improved  data  engine  for  object  detection  in  autonomous  driving,  which  autonomously  identifies  learning  targets,  curates  relevant  data,  auto-labeling  and  continual  learning  with  the  noisy  label,  and  validates  model  improvements,  all  powered  by  vision-language  models  and  large  language  models.These  contributions  not  only  advance  the  theoretical  foundations  of  continual  learning  but  also  address  practical  challenges  in  deploying  adaptive  visual  understanding  systems.  Our  work  demonstrates  the  importance  of  task-specific  considerations  in  continual  learning  and  establishes  a  framework  for  building  self-evolving  computer  vision  systems  that  can  autonomously  adapt  to  the  dynamic  nature  of  real-world  visual  understanding  tasks.
■590    ▼aSchool  code:  0163.
■650  4▼aElectrical  engineering
■650  4▼aComputer  science
■653    ▼aAutomatic  self-improve  data  engine
■653    ▼aAutonomous  driving
■653    ▼aContinual  learning
■653    ▼aContinual  visual  learning
■653    ▼aIncremental  subpopulation  learning
■690    ▼a0544
■690    ▼a0984
■71020▼aNorthwestern  University▼bElectrical  and  Computer  Engineering.
■7730  ▼tDissertations  Abstracts  International▼g86-06B.
■790    ▼a0163
■791    ▼aPh.D.
■792    ▼a2024
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17164670▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

미리보기

내보내기

chatGPT토론

Ai 추천 관련 도서


    신착도서 더보기
    최근 3년간 통계입니다.

    소장정보

    • 예약
    • 소재불명신고
    • 나의폴더
    • 우선정리요청
    • 비도서대출신청
    • 야간 도서대출신청
    소장자료
    등록번호 청구기호 소장처 대출가능여부 대출정보
    TF11898 전자도서 대출가능 마이폴더 부재도서신고 비도서대출신청 야간 도서대출신청

    * 대출중인 자료에 한하여 예약이 가능합니다. 예약을 원하시면 예약버튼을 클릭하십시오.

    해당 도서를 다른 이용자가 함께 대출한 도서

    관련 인기도서

    로그인 후 이용 가능합니다.