Regularizing Image Encoders to Generate Bird's-Eye View Representations for Autonomous Driving Tasks

• Item type: SIG Technical Reports(1)
• 公開日: 2025-01-14
• タイトル: Regularizing Image Encoders to Generate Bird's-Eye View Representations for Autonomous Driving Tasks
• タイトル / 言語: en
• タイトル: Regularizing Image Encoders to Generate Bird's-Eye View Representations for Autonomous Driving Tasks
• 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 Science Tokyo
• 著者所属: Institute of Science Tokyo／National Institute of Informatics
• 著者所属: Denso IT Laboratory
• 著者所属: Institute of Science Tokyo／Denso IT Laboratory
• 著者所属(英): en
• 著者所属(英): Institute of Science Tokyo
• 著者所属(英): en
• 著者所属(英): Institute of Science Tokyo / National Institute of Informatics
• 著者所属(英): en
• 著者所属(英): Denso IT Laboratory
• 著者所属(英): en
• 著者所属(英): Institute of Science Tokyo / Denso IT Laboratory
• 著者名: Qiaoyi, Deng; Satoshi, Ikehata; Yusuke, Sekikawa; Ikuro, Sato
• 著者名(英): Qiaoyi, Deng; Satoshi, Ikehata; Yusuke, Sekikawa; Ikuro, Sato
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Bird's-Eye View (BEV) representations are critical for providing a unified spatial scene understanding to autonomous driving tasks. However, existing methods often struggle with a lack of transformation equivariance. This results in artifacts on BEV feature maps that degrade the performance of downstream tasks. To address this issue, we propose a regularization approach to enhance transformation equivariance through ego-vehicle and dynamic object motion transformations by aligning BEV features in the BEV coordinate system across consecutive frames and introduces a consistency loss to penalize feature misalignment. Experiments on the nuScenes dataset demonstrate that the proposed approach effectively reduces artifacts, stabilizes BEV representations, and improves the reliability of downstream tasks.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Bird's-Eye View (BEV) representations are critical for providing a unified spatial scene understanding to autonomous driving tasks. However, existing methods often struggle with a lack of transformation equivariance. This results in artifacts on BEV feature maps that degrade the performance of downstream tasks. To address this issue, we propose a regularization approach to enhance transformation equivariance through ego-vehicle and dynamic object motion transformations by aligning BEV features in the BEV coordinate system across consecutive frames and introduces a consistency loss to penalize feature misalignment. Experiments on the nuScenes dataset demonstrate that the proposed approach effectively reduces artifacts, stabilizes BEV representations, and improves the reliability of downstream tasks.
• 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 2025-CVIM-240, 号 17, p. 1-4, 発行日 2025-01-14
• 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