Scholarlink

<p>Face recognition has been a pivotal research area in computer vision and biometrics, valued for its non-intrusive and user-friendly characteristics. Recent advancements in deep convolutional neural networks and the availability of large-scale face databases have greatly improved the performance of 2D face recognition. However, its effectiveness can be compromised by variations in illumination, pose, expression, and occlusion, as well as its susceptibility to face presentation attacks. In contrast, 3D data capture preserves geometric shape information, offering inherent robustness to changes in pose and illumination, and greater resilience against face spoofing. The emergence of consumer-grade depth sensors, such as the Microsoft Kinect and Intel RealSense, has further catalyzed the development of 3D face recognition technologies. In this paper, we present the CAS-AIR-3D Face database, a comprehensive large-scale low-quality 3D face database collected using the Intel RealSense SR305 sensor. The database comprises three modalities: color, depth, and near-infrared, and exhibits variations in pose, expression, occlusion, and distance. We also propose a preprocessing pipeline for the CAS-AIR-3D Face database, providing a solution for extracting cropped and aligned faces. Additionally, we introduce multi-modal face recognition benchmarks based on the CAS-AIR-3D Face database. These benchmarks cover image-level fusion, feature-level fusion, and score-level fusion techniques, and explore dual-modal fusion of VIS and depth modalities, dual-modal fusion of NIR and depth modalities, and triple-modal fusion of VIS, NIR, and depth modalities. Furthermore, considering the prevalent use of NIR images in face recognition systems for their robustness to illumination changes, we also evaluate several baseline methods for heterogeneous NIR-VIS face recognition using the CAS-AIR-3D Face database. The database is available for download at <a href="https://github.com/cripac-3dface/CAS-AIR-3D-Face.">https://github.com/cripac-3dface/CAS-AIR-3D-Face.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CAS-AIR-3D: A Large-scale Low-quality Multi-modal Face Database

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Qi Li,
  • Xiaoxiao Dong,
  • Weining Wang,
  • Zhenan Sun,
  • Tieniu Tan,
  • Caifeng Shan

摘要

Face recognition has been a pivotal research area in computer vision and biometrics, valued for its non-intrusive and user-friendly characteristics. Recent advancements in deep convolutional neural networks and the availability of large-scale face databases have greatly improved the performance of 2D face recognition. However, its effectiveness can be compromised by variations in illumination, pose, expression, and occlusion, as well as its susceptibility to face presentation attacks. In contrast, 3D data capture preserves geometric shape information, offering inherent robustness to changes in pose and illumination, and greater resilience against face spoofing. The emergence of consumer-grade depth sensors, such as the Microsoft Kinect and Intel RealSense, has further catalyzed the development of 3D face recognition technologies. In this paper, we present the CAS-AIR-3D Face database, a comprehensive large-scale low-quality 3D face database collected using the Intel RealSense SR305 sensor. The database comprises three modalities: color, depth, and near-infrared, and exhibits variations in pose, expression, occlusion, and distance. We also propose a preprocessing pipeline for the CAS-AIR-3D Face database, providing a solution for extracting cropped and aligned faces. Additionally, we introduce multi-modal face recognition benchmarks based on the CAS-AIR-3D Face database. These benchmarks cover image-level fusion, feature-level fusion, and score-level fusion techniques, and explore dual-modal fusion of VIS and depth modalities, dual-modal fusion of NIR and depth modalities, and triple-modal fusion of VIS, NIR, and depth modalities. Furthermore, considering the prevalent use of NIR images in face recognition systems for their robustness to illumination changes, we also evaluate several baseline methods for heterogeneous NIR-VIS face recognition using the CAS-AIR-3D Face database. The database is available for download at https://github.com/cripac-3dface/CAS-AIR-3D-Face.