• Search Research Projects
  • Search Researchers
  • How to Use
  1. Back to previous page

Unsupervised learning of general-purpose 3D shape feature and its application to analysis of minority 3D shapes

Research Project

Project/Area Number 21K17763
Research Category

Grant-in-Aid for Early-Career Scientists

Allocation TypeMulti-year Fund
Review Section Basic Section 61010:Perceptual information processing-related
Research InstitutionUniversity of Yamanashi

Principal Investigator

FURUYA Takahiko  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
Project Status Completed (Fiscal Year 2023)
Budget Amount *help
¥4,680,000 (Direct Cost: ¥3,600,000、Indirect Cost: ¥1,080,000)
Fiscal Year 2023: ¥1,560,000 (Direct Cost: ¥1,200,000、Indirect Cost: ¥360,000)
Fiscal Year 2022: ¥1,560,000 (Direct Cost: ¥1,200,000、Indirect Cost: ¥360,000)
Fiscal Year 2021: ¥1,560,000 (Direct Cost: ¥1,200,000、Indirect Cost: ¥360,000)
Keywords3次元形状 / 3次元点群 / 教師なし学習 / 自己教師あり学習 / 深層学習 / 形状特徴量
Outline of Research at the Start

3次元(3D)形状データの解析(例:分類,比較,検索,領域分割)では,3D形状特徴量が重要な役割を持つ.近年では深層学習で獲得する形状特徴量が主流となった.しかし,従来の形状特徴量は,学習に用いた特定の種類の3D形状に過剰に適合するために用途が限られる.
本研究は,幅広い形状種に使える「汎用型」の形状特徴量を教師なし深層学習で獲得する.教師なし学習により,ラベルを持たない多量の3D形状を学習に活用できる.さらにスパース符号化技術の導入により特徴量の汎用性を高める.汎用型形状特徴量は特に,製造・医療・建築等の分野に存在する,同種形状サンプルが少ない「少数派」3D形状の解析精度の改善に役立つ.

Outline of Final Research Achievements

In this research, we focused on establishing techniques for unsupervised learning of general-purpose 3D shape features that can be applied to various types of 3D shape data. Specifically, we proposed (1) an unsupervised feature learning algorithm applicable to 3D shape data with diverse shape representations, (2) a novel DNN structure and its loss function for 3D point cloud shape completion, and (3) a new DNN structure and its unsupervised learning method for acquiring rotation-invariant 3D shape features. Through quantitative evaluations, we confirmed that each of the proposed methods demonstrates higher analysis accuracy compared to conventional methods. These research results are expected to improve the analysis accuracy of 3D shapes with a limited number of samples (minority 3D shapes), which have previously relied on handcrafted 3D shape features.

Academic Significance and Societal Importance of the Research Achievements

本研究の学術的意義は,3D形状データの種類を問わず使える汎用型形状特徴量を教師なし深層学習で獲得する手法を提案したことである.これにより,ラベル付きデータが少ない3D形状の解析においても,ある程度高精度な特徴量を利用できる.
社会的意義として,製造業,建築,医療,娯楽作品制作といった3D形状データを活用する様々な分野において,3D形状解析の自動化・高精度化が期待できる.特に,ラベル付きデータの収集が困難な分野では本研究で提案した教師なし学習手法が有効に活用できると考えられる.さらには,各分野における3D形状データの活用が促進され,生産性の向上や新たな価値創造につながることが期待できる.

Report

(4 results)
  • 2023 Annual Research Report   Final Research Report ( PDF )
  • 2022 Research-status Report
  • 2021 Research-status Report
  • Research Products

    (11 results)

All 2024 2023 2022 2021 Other

All Int'l Joint Research (2 results) Journal Article (5 results) (of which Int'l Joint Research: 1 results,  Peer Reviewed: 4 results,  Open Access: 5 results) Presentation (4 results) (of which Invited: 3 results)

  • [Int'l Joint Research] Hangzhou Dianzi University(中国)

    • Related Report
      2023 Annual Research Report
  • [Int'l Joint Research] Hangzhou Dianzi University(中国)

    • Related Report
      2022 Research-status Report
  • [Journal Article] Self-supervised Learning of Rotation-invariant 3D Point Set Features using Transformer and its Self-distillation2024

    • Author(s)
      Takahiko Furuya, Zhoujie Chen, Ryutarou Ohbuchi, Zhenzhong Kuang
    • Journal Title

      Computer Vision and Image Understanding

      Volume: 244 Pages: 104025-104025

    • DOI

      10.1016/j.cviu.2024.104025

    • Related Report
      2023 Annual Research Report
    • Peer Reviewed / Open Access
  • [Journal Article] 自己注意機構を用いた,3次元点群形状の回転不変な解析2023

    • Author(s)
      刈込喜大,古屋貴彦,大渕竜太郎
    • Journal Title

      画像電子学会誌

      Volume: 52 Pages: 516-526

    • Related Report
      2023 Annual Research Report
    • Peer Reviewed / Open Access
  • [Journal Article] DeepDiffusion: Unsupervised Learning of Retrieval-Adapted Representations via Diffusion-Based Ranking on Latent Feature Manifold2022

    • Author(s)
      Takahiko Furuya, Ryutarou Ohbuchi
    • Journal Title

      IEEE Access

      Volume: 10 Pages: 116287-116301

    • DOI

      10.1109/access.2022.3218909

    • Related Report
      2022 Research-status Report
    • Peer Reviewed / Open Access
  • [Journal Article] Hyperplane patch mixing-and-folding decoder and weighted chamfer distance loss for 3D point set reconstruction2022

    • Author(s)
      Takahiko Furuya, Wujie Liu, Ryutarou Ohbuchi, Zhenzhong Kuang
    • Journal Title

      The Visual Computer

      Volume: 0 Issue: 10 Pages: 1-18

    • DOI

      10.1007/s00371-022-02652-6

    • Related Report
      2022 Research-status Report
    • Peer Reviewed / Open Access / Int'l Joint Research
  • [Journal Article] DeepDiffusion: Unsupervised Learning of Retrieval-adapted Representations via Diffusion-based Ranking on Latent Feature Manifold2021

    • Author(s)
      Takahiko Furuya, Ryutarou Ohbuchi
    • Journal Title

      arXiv preprint

      Volume: 2112.07082

    • Related Report
      2021 Research-status Report
    • Open Access
  • [Presentation] 深層学習を用いた点群処理(中/上級)2023

    • Author(s)
      古屋貴彦
    • Organizer
      精密工学会,大規模環境の3次元計測と認識・モデル化技術専門委員会,技術講習会
    • Related Report
      2023 Annual Research Report
    • Invited
  • [Presentation] 深層学習を用いた点群処理(中/上級)2022

    • Author(s)
      古屋 貴彦
    • Organizer
      大規模環境の3次元計測と認識・モデル化技術専門委員会,点群処理基礎技術講習会
    • Related Report
      2022 Research-status Report
    • Invited
  • [Presentation] 自己注意機構を用いた3次元点群形状の解析2021

    • Author(s)
      刈込 喜大
    • Organizer
      画像の認識・理解シンポジウム(MIRU)2021
    • Related Report
      2021 Research-status Report
  • [Presentation] 深層学習を用いた点群処理(中級)・3次元点群深層学習の解説と実践2021

    • Author(s)
      古屋 貴彦
    • Organizer
      大規模環境の3次元計測と認識・モデル化技術専門委員会 技術講習会
    • Related Report
      2021 Research-status Report
    • Invited

URL: 

Published: 2021-04-28   Modified: 2025-01-30  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi