Learning from Synthetic Shadows

• Item type: SIG Technical Reports(1)
• 公開日: 2021-02-25
• タイトル: Learning from Synthetic Shadows
• タイトル / 言語: en
• タイトル: Learning from Synthetic Shadows
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: セッション5-1
• 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: Department of Information and Communication Engineering, The University of Tokyo
• 著者所属: Department of Information and Communication Engineering, The University of Tokyo
• 著者所属(英): en
• 著者所属(英): Department of Information and Communication Engineering, The University of Tokyo
• 著者所属(英): en
• 著者所属(英): Department of Information and Communication Engineering, The University of Tokyo
• 著者名: Naoto, Inoue; Toshihiko, Yamasaki
• 著者名(英): Naoto, Inoue; Toshihiko, Yamasaki
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Shadow removal is essential for some downstream tasks in computer vision and computer graphics. Recent shadow removal approaches all train convolutional neural networks (CNN) on real paired shadow/shadow-free/mask image datasets. However, obtaining a large-scale, diverse, and accurate dataset has been a primary challenge. It limits the generalization performance of the learned models on shadow images with unseen shapes/intensities. We present SynShadow, a novel large-scale synthetic shadow/shadow-free/matte image triplets dataset and a pipeline to synthesize it to tackle this challenge. We extend a physically-grounded shadow illumination model and synthesize a shadow image given an arbitrary combination of a shadow-free image, a matte image, and shadow attenuation parameters. This pipeline enables us to sample a countless number of the triplets. SynShadow offers a dataset with high fidelity and diversity. We demonstrate that shadow removal models trained on SynShadow perform favorably in removing shadows with various shapes and intensities. Furthermore, we show that simply fine-tuning from a SynShadow-pre-trained model improves existing shadow detection and removal models.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Shadow removal is essential for some downstream tasks in computer vision and computer graphics. Recent shadow removal approaches all train convolutional neural networks (CNN) on real paired shadow/shadow-free/mask image datasets. However, obtaining a large-scale, diverse, and accurate dataset has been a primary challenge. It limits the generalization performance of the learned models on shadow images with unseen shapes/intensities. We present SynShadow, a novel large-scale synthetic shadow/shadow-free/matte image triplets dataset and a pipeline to synthesize it to tackle this challenge. We extend a physically-grounded shadow illumination model and synthesize a shadow image given an arbitrary combination of a shadow-free image, a matte image, and shadow attenuation parameters. This pipeline enables us to sample a countless number of the triplets. SynShadow offers a dataset with high fidelity and diversity. We demonstrate that shadow removal models trained on SynShadow perform favorably in removing shadows with various shapes and intensities. Furthermore, we show that simply fine-tuning from a SynShadow-pre-trained model improves existing shadow detection and removal models.
• 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 2021-CVIM-225, 号 38, p. 1-6, 発行日 2021-02-25
• 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