Full metadata record
DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | 최용석 | - |
dc.contributor.author | 장두수 | - |
dc.date.accessioned | 2021-02-24T16:01:03Z | - |
dc.date.available | 2021-02-24T16:01:03Z | - |
dc.date.issued | 2021. 2 | - |
dc.identifier.uri | https://repository.hanyang.ac.kr/handle/20.500.11754/158927 | - |
dc.identifier.uri | http://hanyang.dcollection.net/common/orgView/200000485721 | en_US |
dc.description.abstract | Zero-shot recognition (ZSR) aims to learn a visual classifier for categories with the absence of training samples. The focus in most traditional ZSR models is to transfer semantic knowledge of familiar categories to represent unfamiliar ones only with the visual appearance of an unseen object. To transfer knowledge, it is assumed that visually similar objects tend to also be semantically similar. However, the correlation between visual and semantic characteristics is not always assured, which induces a limitation in accurate and wide-covered predictions. To alleviate this limitation, we consider not only the visual information but contexts to enhance cognitive ability in a multi-object scene. In this paper, we propose a novel method, contextual inference, that exploits the surrounding information obtained with similarity measures between an unseen object and its surrounding objects by using external resources such as knowledge graphs and semantic embedding spaces. Moreover, with the intuition that close contexts involve more related associations, distance-weighting is applied to each of the surrounding information by a newly defined distance calculation formula. We integrate the contextual inference into the traditional ZSR models to calibrate their visual predictions. We perform extensive experiments for comparative evaluations on two different datasets: ImageNet categories and Visual Genome categories containing small and large amounts of target categories, respectively. Various quantitative experiments with accuracy metrics and qualitative experiments with exemplary prediction results are conducted for validation. The experimental results demonstrate the effectiveness of our method with significant enhancements in performances. Consequently, we expect this dissertation plays a key role in investigating context-aware ZSR that is yet rarely studied. | - |
dc.publisher | 한양대학교 | - |
dc.title | Similarity-based Contextual Zero-Shot Recognition via Distance-Weighted Calibration | - |
dc.title.alternative | 거리-가중 보정을 통한 유사도 기반 문맥적 제로샷 인식 | - |
dc.type | Theses | - |
dc.contributor.googleauthor | Doo Soo Chang | - |
dc.contributor.alternativeauthor | 장두수 | - |
dc.sector.campus | S | - |
dc.sector.daehak | 대학원 | - |
dc.sector.department | 컴퓨터·소프트웨어학과 | - |
dc.description.degree | Doctor | - |
dc.contributor.affiliation | 인공지능 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.