Non-learning depth completion with continuous and binary anisotropic diffusion tensor

• Item type: SIG Technical Reports(1)
• 公開日: 2024-01-18
• タイトル: Non-learning depth completion with continuous and binary anisotropic diffusion tensor
• タイトル / 言語: en
• タイトル: Non-learning depth completion with continuous and binary anisotropic diffusion tensor
• 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: 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
• 著者所属(英): en
• 著者所属(英): Institute of Industrial Science, The University of Tokyo
• 著者名: Guan, Ziang; Yasuhiro, Yao; Ryoichi, Ishikawa; Takeshi, Oishi
• 著者名(英): Guan, Ziang; Yasuhiro, Yao; Ryoichi, Ishikawa; Takeshi, Oishi
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: We present a high-performance non-learning depth completion algorithm aimed at swiftly and accurately enhancing sparse lidar-generated depth maps to denser versions. The algorithm's distinctive feature lies in its capacity to guide the depth completion process based solely on scene images corresponding to the depth maps. Despite the established effectiveness of learning-based depth completion methods, particularly those employing deep neural networks, the challenge lies in acquiring a substantial amount of annotated data for model training, often challenging in practical scenarios. Our paper presents a novel framework that leverages geometric scene information for depth completion. It encompasses three key components: an algorithm for mis-projection point elimination, an image-guided interpolation algorithm, and a depth map refinement process utilizing Diffusion Tensor. Through comprehensive testing, our method outperforms existing non-learning depth completion techniques in overall error metrics and shows superior performance in preserving object edges and surface details in complex scenes. These results highlight the effectiveness and practicality of our approach in real-world applications.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: We present a high-performance non-learning depth completion algorithm aimed at swiftly and accurately enhancing sparse lidar-generated depth maps to denser versions. The algorithm's distinctive feature lies in its capacity to guide the depth completion process based solely on scene images corresponding to the depth maps. Despite the established effectiveness of learning-based depth completion methods, particularly those employing deep neural networks, the challenge lies in acquiring a substantial amount of annotated data for model training, often challenging in practical scenarios. Our paper presents a novel framework that leverages geometric scene information for depth completion. It encompasses three key components: an algorithm for mis-projection point elimination, an image-guided interpolation algorithm, and a depth map refinement process utilizing Diffusion Tensor. Through comprehensive testing, our method outperforms existing non-learning depth completion techniques in overall error metrics and shows superior performance in preserving object edges and surface details in complex scenes. These results highlight the effectiveness and practicality of our approach in real-world 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 2024-CVIM-236, 号 15, 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