SWIN-RIND: Edge Detection for Reflectance, Illumination, Normal and Depth Discontinuity with Swin Transformer

• Item type: SIG Technical Reports(1)
• 公開日: 2024-01-18
• タイトル: SWIN-RIND: Edge Detection for Reflectance, Illumination, Normal and Depth Discontinuity with Swin Transformer
• タイトル / 言語: en
• タイトル: SWIN-RIND: Edge Detection for Reflectance, Illumination, Normal and Depth Discontinuity with Swin Transformer
• 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: Institute of Industrial Science, The University of Tokyo
• 著者所属: Institute of Industrial Science, The University of Tokyo
• 著者所属: Institute of Industrial Science, The University of Tokyo
• 著者所属(英): en
• 著者所属(英): Institute of Industrial Science, The University of Tokyo
• 著者所属(英): en
• 著者所属(英): Institute of Industrial Science, The University of Tokyo
• 著者所属(英): en
• 著者所属(英): Institute of Industrial Science, The University of Tokyo
• 著者名: Lun, Miao; Ryoichi, Ishikawa; Takeshi, Oishi
• 著者名(英): Lun, Miao; Ryoichi, Ishikawa; Takeshi, Oishi
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Edges are caused by the discontinuities in Surface-Reflectance, Illumination, Surface-Normal, and Depth (RIND). However, despite general edge detection being studied for decades, research on specific edges has not been extensively explored. In this work, we propose a transformer-based approach called SWIN-RIND that can detect the four types of edges from a single image. Recently, attention-based approaches have performed well in general edge detection and can be expected to work for RIND edges as well. Our model uses Swin Transformer as an encoder and a top-down and bottom-up multi-level feature aggregation block as a decoder. The encoder extracts cues at different levels, which the decoder integrates into shared features containing rich contextual information. We then predict each specific edge type through independent decision heads. To train and evaluate our model, we use a public benchmark called BSDS-RIND, which is based on BSDS (Berkeley Segmentation Data Set) and contains annotations of four types of edges. In our experiments, we confirmed that SWIN-RIND outperforms state-of-the-art methods.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Edges are caused by the discontinuities in Surface-Reflectance, Illumination, Surface-Normal, and Depth (RIND). However, despite general edge detection being studied for decades, research on specific edges has not been extensively explored. In this work, we propose a transformer-based approach called SWIN-RIND that can detect the four types of edges from a single image. Recently, attention-based approaches have performed well in general edge detection and can be expected to work for RIND edges as well. Our model uses Swin Transformer as an encoder and a top-down and bottom-up multi-level feature aggregation block as a decoder. The encoder extracts cues at different levels, which the decoder integrates into shared features containing rich contextual information. We then predict each specific edge type through independent decision heads. To train and evaluate our model, we use a public benchmark called BSDS-RIND, which is based on BSDS (Berkeley Segmentation Data Set) and contains annotations of four types of edges. In our experiments, we confirmed that SWIN-RIND outperforms state-of-the-art methods.
• 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 21, p. 1-8, 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