キクチ　マサユキ   MASAYUKI KIKUCHI   菊池　眞之    所属   コンピュータサイエンス学部 コンピュータサイエンス学科    職種   准教授
言語種別	英語
発行・発表の年月	2022/12
形態種別	国際会議論文
査読	査読あり
標題	A model of 3D surface ownership assignment
執筆形態	共著
掲載誌名	Perception (The 44th European Conference on Visual Perception (ECVP) 2022)
掲載区分	国外
出版社・発行元	SAGE journals
巻・号・頁	51(1 suppl)
総ページ数	1
担当区分	筆頭著者
著者・共著者	Masayuki Kikuchi, Shunta Ishikawa
概要	It has become clear that the mammalian visual system has the ability to assign border-ownership for 2D object images. On the other hand, since the 3D world is usually viewed binocularly, borders of objects correspond 3D surfaces instead of 2D contours. Therefore, the visual system needs to determine the side of object region at each local surface area. Alike 2D cases, 3D objects have both convex and concave parts, and the proportion of convex parts always exceed concave parts. With such property in mind, we constructed a neural network model assigning ownership of 3D surfaces by extending the previous 2D border-ownership model proposed by one of us. Dimension of the input space is 3D assuming the stage after binocular correspondence problem is solved. At each surface point, a pair of surface ownership neurons are prepared and connected mutually with the same ownership polarity within a local spatial range, contributing to smooth the responses. In addition, mutual inhibition is equipped between each pair of antagonistic ownership neurons. Absolute values of curvature are given as initial value of neurons for inner side of the local ridges of surface, irrespective of global inside/outside. After iterative calculation, positive responses remain only on the neurons corresponding global interior of the surface. We tested the model on a computer. For 3D objects with random convex and concave structures, the model detected successfully global inner side.
外部リンクURL	https://journals.sagepub.com/doi/epub/10.1177/03010066221141167