Outlier Detection for Robust Parameter Estimation Against Multi-modeled/structured Data

• Item type: SIG Technical Reports(1)
• 公開日: 2010-05-20
• タイトル: Outlier Detection for Robust Parameter Estimation Against Multi-modeled/structured Data
• タイトル / 言語: en
• タイトル: Outlier Detection for Robust Parameter Estimation Against Multi-modeled/structured Data
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: Ｄ論セッション２
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Osaka University
• 著者所属: Kyushu University
• 著者所属: Advanced Industrial Science and Technology
• 著者所属: Osaka University
• 著者所属: Osaka Institute of Technology
• 著者所属: Osaka University
• 著者所属(英): en
• 著者所属(英): Osaka University
• 著者所属(英): en
• 著者所属(英): Kyushu University
• 著者所属(英): en
• 著者所属(英): Advanced Industrial Science and Technology
• 著者所属(英): en
• 著者所属(英): Osaka University
• 著者所属(英): en
• 著者所属(英): Osaka Institute of Technology
• 著者所属(英): en
• 著者所属(英): Osaka University
• 著者名: NgoTrungThanh; Hajime, Nagahara; Ryusuke, Sagawa; Yasuhiro, Mukaigawa; Masahiko, Yachida; Yasushi, Yagi
• 著者名(英): Ngo, TrungThanh; Hajime, Nagahara; Ryusuke, Sagawa; Yasuhiro, Mukaigawa; Masahiko, Yachida; Yasushi, Yagi
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Model parameter estimation and automatic outlier detection is a fundamental and important problem in computer vision. Vision data is noisy and usually contains multiple structures, models of interest. RANSAC has been proven to be the most popular and effective solution for such problem, but it requires some user-defined threshold to discriminate inliers/outliers. It is then improved by the adaptive-scale robust estimators, which do not require the user-defined threshold and detect inliers automatically. However, there still remains some problem. The problem is that these adaptive-scale robust estimators do not focus on the accurate inlier detection. In this paper, we propose several adaptive-scale robust estimators which can detect inliers accurately. There are two reasons for the idea of accurate inlier detection. First, if a robust estimator detects inliers better, then the robustness of the estimation can be improved. Second, in many real applications such as motion segmentation and range image segmentation, if the inlier detection is not very well, then a structure can be broken into smaller structures, an under-segmentation problem, or united with the other structures, an oversegmentation problem. In the experiments, various analytic simulations in many aspects have shown the advantage of the proposed robust estimators compared to several latest robust estimators. The real experiments were also performed to prove the validation of the proposed estimators in real applications.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Model parameter estimation and automatic outlier detection is a fundamental and important problem in computer vision. Vision data is noisy and usually contains multiple structures, models of interest. RANSAC has been proven to be the most popular and effective solution for such problem, but it requires some user-defined threshold to discriminate inliers/outliers. It is then improved by the adaptive-scale robust estimators, which do not require the user-defined threshold and detect inliers automatically. However, there still remains some problem. The problem is that these adaptive-scale robust estimators do not focus on the accurate inlier detection. In this paper, we propose several adaptive-scale robust estimators which can detect inliers accurately. There are two reasons for the idea of accurate inlier detection. First, if a robust estimator detects inliers better, then the robustness of the estimation can be improved. Second, in many real applications such as motion segmentation and range image segmentation, if the inlier detection is not very well, then a structure can be broken into smaller structures, an under-segmentation problem, or united with the other structures, an oversegmentation problem. In the experiments, various analytic simulations in many aspects have shown the advantage of the proposed robust estimators compared to several latest robust estimators. The real experiments were also performed to prove the validation of the proposed estimators in real applications.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11131797
• 書誌情報: 研究報告コンピュータビジョンとイメージメディア（CVIM）; 巻 2010-CVIM-172, 号 38, p. 1-17, 発行日 2010-05-20
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会