자료유형  

 학위논문

Control Number  

0017161407

International Standard Book Number  

9798382842929

Dewey Decimal Classification Number  

004

Main Entry-Personal Name  

Phoo, Cheng Perng.

Publication, Distribution, etc. (Imprint  

[S.l.] : Cornell University., 2024

Publication, Distribution, etc. (Imprint  

Ann Arbor : ProQuest Dissertations & Theses, 2024

Physical Description  

350 p.

General Note  

Source: Dissertations Abstracts International, Volume: 85-12, Section: A.

General Note  

Advisor: Hariharan, Bharath.

Dissertation Note  

Thesis (Ph.D.)--Cornell University, 2024.

Summary, Etc.  

요약For the past decades, we have observed tremendous success in developing perception models - computational models that could perceive our world through images, videos, LiDAR point clouds, and so on. Currently, we have perception models that can recognize thousands of concepts commonly seen on the Internet. The ability of these models to recognize concepts is undeniably impressive, but their successes are only limited to concepts or data modalities (e.g. images) commonly seen on the Internet.Beyond applications in the Internet domain such as remote sensing or medical imagery, perception models have yet to show their prowess. The key challenge in building perception models beyond Internet applications is the requirement of extensive expert involvement. Training performant perception models in these domains often requires non-trivial involvement from experts, especially during the data collection process.In this dissertation, we investigate how we could reduce experts' burden when developing perception models. Specifically, we will focus on the angle of label efficiency, i.e., developing perception models that could be trained with fewer annotations. We will present two broad categories of approaches. The first category relies on minimal assumptions and could be applied to various problem domains; along this vein, we will examine how we could leverage pre-trained models, unlabeled data, and coarsely-labeled data to enhance label efficiency. The second category leverages domain knowledge to enhance label efficiency. For this category of approaches, we will look at two specific domains: autonomous driving and remote sensing. We will investigate how repeated traversals of the same location could be used to improve perception models for self-driving vehicles and how ground images could be used to train vision-language models for remote sensing without any textual annotations. We will end this dissertation with a brief discussion of how we could further reduce experts' burden when developing perception models, enabling broader success of perception models beyond Internet applications.

Subject Added Entry-Topical Term  

Computer science.

Subject Added Entry-Topical Term  

Web studies.

Index Term-Uncontrolled  

Computer vision

Index Term-Uncontrolled  

Fewer annotations

Index Term-Uncontrolled  

Machine perception

Index Term-Uncontrolled  

Internet applications

Index Term-Uncontrolled  

Data collection

Added Entry-Corporate Name  

Cornell University Computer Science

Host Item Entry  

Dissertations Abstracts International. 85-12A.

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:658307