서브메뉴
검색
Representation Learning for Active Perception
Representation Learning for Active Perception
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104841
- ISBN
- 9798293892471
- DDC
- 621.3
- 저자명
- Quenum, Jerome.
- 서명/저자
- Representation Learning for Active Perception
- 발행사항
- [Sl] : University of California, Berkeley, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 156 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-04, Section: B.
- 주기사항
- Advisor: Darrell, Trevor;Malik, Jitendra.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2025.
- 초록/해제
- 요약Active perception driven by representation learning lies at the intersection of computer vision and robotics, powered by advanced Artificial Intelligence (AI) algorithms. It has for goal to develop systems that can actively gather information from their environment by learning useful representations to enhance both their perception and decision-making capabilities. This dissertation explores various facets of this topic, with a particular focus on how advanced representation-learning techniques can significantly improve the performance and utility of such systems.We begin by examining the role of precision in AI-driven systems through classical segmentation tasks in Ultra-High-Resolution (UHR) images within the context of battery design. A novel transformer-based network, TransforCNN, is introduced to segment dendrites in Lithium Metal Battery (LMB) 3D X-ray computed tomography (XCT) volumes. This approach significantly improves the accuracy of segmenting critical structures, which is essential for developing more efficient and reliable batteries. The precision gained in this segmentation problem through our proposed representation learning model not only enhances the visual understanding of battery components but also lays the groundwork for more informed and effective design strategies.Building on this foundation, we transition to addressing the challenge of processing Ultra-High-Resolution (UHR) images, which is important for applications requiring both speed and accuracy. We explore this in the context of barcode detection in UHR images, where a new detection pipeline is proposed, integrating a modified Region Proposal Network (RPN) with a proposed segmentation network named Y-Net. This study demonstrates how representation learning can be optimized to reduce latency while maintaining high accuracy, showcasing the potential of AI to handle complex visual tasks in real time.We then examine scalability when it comes to the integration of representation learning for active perception in robotics, where speed, accuracy, adaptability, and robustness are key. Using the Open X Embodiment (OXE) dataset, we investigate a cross-robot self-supervised sensorimotor pre-training approach through RPTx, a multimodal model that does not leverage language instructions and does not use auto-regressive methods. This approach provided valuable insights into the robustness and adaptability issues encountered by robotic systems. Here, we conduct reliability assessments, an examination of negative transfer instances where models struggle to adapt, and we speculate on future directions for overcoming these difficulties. Additionally, we observe that the use of multimodal models that leverage large language models for instruction tuning and that employ auto-regressive techniques led to more stable and promising outcomes with LLARVA, a vision-action instruction tuning paradigm that enhances Robot Learning. These findings demonstrate how learning good representations can enhance robotic systems' ability to interpret and respond to complex environmental cues.Finally, we explore the challenge of extending segmentation models beyond fixed object categories to support reasoning over visual scenes. While current segmentation models work well when objects are clearly defined, they struggle with complex user queries that refer to multiple or implicit objects. Recent work in reasoning segmentation, which generates segmentation masks from natural language input, shows that vision-language models (VLMs) can help address this. However, our experiments show that existing models perform poorly on remote-sensing images, which often contain dense and varied content. To address this, we introduce LISAt, a vision-language model designed to describe remote-sensing images, answer questions about them, and segment objects based on user queries. LISAt is trained on GRES, a dataset with 27,615 annotations across 9,205 images, and PreGRES, a multimodal dataset with over one million question-answer pairs. LISAt outperforms geospatial models like RS-GPT4V by 10.04% in BLEU-4 for image description and improves on open-domain models in reasoning segmentation by 143.36% in generalized Intersection over Union (gIoU). These results show how learning strong multimodal representations can improve scene understanding, especially in the geospatial domain.
- 일반주제명
- Electrical engineering
- 일반주제명
- Computer science
- 일반주제명
- Information technology
- 기타저자
- University of California, Berkeley Electrical Engineering & Computer Sciences
- 기본자료저록
- Dissertations Abstracts International. 87-04B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017359144
■00520260202104841
■006m o d
■007cr#unu||||||||
■020 ▼a9798293892471
■035 ▼a(MiAaPQ)AAI32172894
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621.3
■1001 ▼aQuenum, Jerome.
■24510▼aRepresentation Learning for Active Perception
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a156 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-04, Section: B.
■500 ▼aAdvisor: Darrell, Trevor;Malik, Jitendra.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2025.
■520 ▼aActive perception driven by representation learning lies at the intersection of computer vision and robotics, powered by advanced Artificial Intelligence (AI) algorithms. It has for goal to develop systems that can actively gather information from their environment by learning useful representations to enhance both their perception and decision-making capabilities. This dissertation explores various facets of this topic, with a particular focus on how advanced representation-learning techniques can significantly improve the performance and utility of such systems.We begin by examining the role of precision in AI-driven systems through classical segmentation tasks in Ultra-High-Resolution (UHR) images within the context of battery design. A novel transformer-based network, TransforCNN, is introduced to segment dendrites in Lithium Metal Battery (LMB) 3D X-ray computed tomography (XCT) volumes. This approach significantly improves the accuracy of segmenting critical structures, which is essential for developing more efficient and reliable batteries. The precision gained in this segmentation problem through our proposed representation learning model not only enhances the visual understanding of battery components but also lays the groundwork for more informed and effective design strategies.Building on this foundation, we transition to addressing the challenge of processing Ultra-High-Resolution (UHR) images, which is important for applications requiring both speed and accuracy. We explore this in the context of barcode detection in UHR images, where a new detection pipeline is proposed, integrating a modified Region Proposal Network (RPN) with a proposed segmentation network named Y-Net. This study demonstrates how representation learning can be optimized to reduce latency while maintaining high accuracy, showcasing the potential of AI to handle complex visual tasks in real time.We then examine scalability when it comes to the integration of representation learning for active perception in robotics, where speed, accuracy, adaptability, and robustness are key. Using the Open X Embodiment (OXE) dataset, we investigate a cross-robot self-supervised sensorimotor pre-training approach through RPTx, a multimodal model that does not leverage language instructions and does not use auto-regressive methods. This approach provided valuable insights into the robustness and adaptability issues encountered by robotic systems. Here, we conduct reliability assessments, an examination of negative transfer instances where models struggle to adapt, and we speculate on future directions for overcoming these difficulties. Additionally, we observe that the use of multimodal models that leverage large language models for instruction tuning and that employ auto-regressive techniques led to more stable and promising outcomes with LLARVA, a vision-action instruction tuning paradigm that enhances Robot Learning. These findings demonstrate how learning good representations can enhance robotic systems' ability to interpret and respond to complex environmental cues.Finally, we explore the challenge of extending segmentation models beyond fixed object categories to support reasoning over visual scenes. While current segmentation models work well when objects are clearly defined, they struggle with complex user queries that refer to multiple or implicit objects. Recent work in reasoning segmentation, which generates segmentation masks from natural language input, shows that vision-language models (VLMs) can help address this. However, our experiments show that existing models perform poorly on remote-sensing images, which often contain dense and varied content. To address this, we introduce LISAt, a vision-language model designed to describe remote-sensing images, answer questions about them, and segment objects based on user queries. LISAt is trained on GRES, a dataset with 27,615 annotations across 9,205 images, and PreGRES, a multimodal dataset with over one million question-answer pairs. LISAt outperforms geospatial models like RS-GPT4V by 10.04% in BLEU-4 for image description and improves on open-domain models in reasoning segmentation by 143.36% in generalized Intersection over Union (gIoU). These results show how learning strong multimodal representations can improve scene understanding, especially in the geospatial domain.
■590 ▼aSchool code: 0028.
■650 4▼aElectrical engineering
■650 4▼aComputer science
■650 4▼aInformation technology
■653 ▼aActive perception
■653 ▼aLatency optimization
■653 ▼aMultimodal representations
■653 ▼aObject detection and segmentation
■653 ▼aReasoning segmentation
■653 ▼aRepresentation learning
■690 ▼a0800
■690 ▼a0544
■690 ▼a0984
■690 ▼a0489
■71020▼aUniversity of California, Berkeley▼bElectrical Engineering & Computer Sciences.
■7730 ▼tDissertations Abstracts International▼g87-04B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359144▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


