A Study on Fixed Orthogonal Prototype Classifier for Semantic Segmentation

• Item type: SIG Technical Reports(1)
• 公開日: 2024-01-18
• タイトル: A Study on Fixed Orthogonal Prototype Classifier for Semantic Segmentation
• タイトル / 言語: en
• タイトル: A Study on Fixed Orthogonal Prototype Classifier for Semantic Segmentation
• 言語: eng
• 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: Faculty of Information, Nagoya University
• 著者所属: Faculty of Information, Nagoya University
• 著者所属: Faculty of Information, Nagoya University
• 著者所属: Faculty of Information, Nagoya University
• 著者所属(英): en
• 著者所属(英): Faculty of Information, Nagoya University
• 著者所属(英): en
• 著者所属(英): Faculty of Information, Nagoya University
• 著者所属(英): en
• 著者所属(英): Faculty of Information, Nagoya University
• 著者所属(英): en
• 著者所属(英): Faculty of Information, Nagoya University
• 著者名: Jialei, Chen; Daisuke, Deguchi; Chenkai, Zhang; Xu, Zheng; Hiroshi, Murase
• 著者名(英): Jialei, Chen; Daisuke, Deguchi; Chenkai, Zhang; Xu, Zheng; Hiroshi, Murase
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Semantic segmentation models consist of an image encoder and a learnable classifier to do this task. Though various image encoders are proposed, few works concentrate on designing classifiers. In this paper, we observe that frozen orthogonal prototypes can work even better than learnable ones with the help of constraints for shaping feature spaces. Therefore, we also propose a loss function to better shape the feature space. Experiments on the ADE20k dataset show impressive results.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Semantic segmentation models consist of an image encoder and a learnable classifier to do this task. Though various image encoders are proposed, few works concentrate on designing classifiers. In this paper, we observe that frozen orthogonal prototypes can work even better than learnable ones with the help of constraints for shaping feature spaces. Therefore, we also propose a loss function to better shape the feature space. Experiments on the ADE20k dataset show impressive results.
• 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 2024-CVIM-236, 号 29, p. 1-2, 発行日 2024-01-18
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8701
• 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