2023 Fiscal Year Final Research Report
Part-based 3D shape retrieval using multi-modal query
| Project/Area Number |
21K11903
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Multi-year Fund |
|---|
| Section | 一般 |
|---|
| Review Section |
Basic Section 60080:Database-related
|
|---|
| Research Institution | University of Yamanashi |
|---|
Principal Investigator |
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
古屋 貴彦 山梨大学, 大学院総合研究部, 准教授 (00770835)
|
|---|
| Project Period (FY) |
2021-04-01 – 2024-03-31
|
| Keywords | 3D shape analysis / multimodal retrieval / computer vision / deep learning / self-supervised learning / feature representation / transformer |
|---|
| Outline of Final Research Achievements |
3D shape data is utilized in diverse fields such as industrial product design, visual content creation, infrastructure (e.g., roads) maintenance, and medical image diagnosis. The goal of this research is an integrated, exploratory, and iterative approach toward effective and efficient part/whole shape retrieval for effective and efficient management of 3D shape data. Toward this goal, new methods were proposed and evaluated to address the following two sub-problems: (1) 3D rotation-invariant 3D shape feature extraction using unsupervised or self-supervised learning, and (2) 3D shape reconstruction method robust against input noise, missing parts, or locally variable sampling resolution.
|
| Free Research Field |
3次元形状データの検索と解析
|
| Academic Significance and Societal Importance of the Research Achievements |
社会的意義は，急速に利用が広がる3D形状データの解析，検索，識別などに欠かせない形状特徴を抽出するより良い手法を提案したことである．学術的な注目点は，（1）教師無し学習ないし自己教師あり学習に基づく，(2) 3次元回転に対し一定の不変性を持つ，である．ラベル付き3D形状データはその数が少なく，また多様性も限られる．そのため，特徴の学習においてラベルが不要の教師無し学習ないし自己教師あり学習が必須である．また，3D形状特徴には3軸周りの回転への不変性を要求される場合が多い．我々が提案した[ XX]は，世界で初めて，回転不変な3D点群形状特徴の自己教師あり学習による獲得に成功した．
|
