Learning Context-Aware Representations for Semantic Segmentation : 의미론적 영상 분할을 위한 맥락 인식 기반 표현 학습

Cited 0 time in Web of Science Cited 0 time in Scopus


공과대학 전기·컴퓨터공학부
Issue Date
서울대학교 대학원
Computer VisionObject RecognitionSemantic SegmentationContext
학위논문 (박사)-- 서울대학교 대학원 : 전기·컴퓨터공학부, 2017. 2. 이경무.
Semantic segmentation, segmenting all the objects and identifying their categories, is a fundamental and important problem in computer vision. Traditional approaches to semantic segmentation are based on two main elements: visual appearance features and semantic context. Visual appearance features such as color, edge, shape and so on, are a primary source of information for reasoning the objects in an image. However, image data are sometimes unable to fully capture diversity in the object classes, since the appearance of the objects presented in real world scenes is affected by imaging conditions such as illumination, texture, occlusion, and viewpoint. Therefore, semantic context, obtained from not only the presence but also the location of other objects, can help to disambiguate the visual appearance in semantic segmentation tasks. The modern contextualized semantic segmentation systems have successfully improved segmentation performance by refining inconsistently labeled pixels via modeling of contextual interactions. However, they considered semantic context and visual appearance features independently due to the absence of the suitable representation model. Motivated by this issue, this dissertation proposes a novel framework for learning semantic context-aware representations in which appearance features is enhanced and enriched by semantic context and vice versa.
The first part of the dissertation will be devoted to semantic context-aware appearance modeling for semantic segmentation. Adaptive context aggregation network is studied to capture semantic context adequately while multiple steps of reasoning. Secondly, semantic context will be reinforced by utilizing visual appearance. Graph and example-based context model is presented for estimating contextual relationships according to the visual appearance of objects. Finally, we propose a Multiscale Conditional Random Fields (CRFs), for integrating context-aware appearance and appearance-aware semantic context to produce accurate segmentations. Experimental evaluations show the effectiveness of the proposed context-aware representations on various challenging datasets.
Files in This Item:
Appears in Collections:
College of Engineering/Engineering Practice School (공과대학/대학원)Dept. of Electrical and Computer Engineering (전기·정보공학부)Theses (Ph.D. / Sc.D._전기·정보공학부)
  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.