研究目的
To reconstruct higher-resolution 3D face models from low-resolution depth frames acquired with a consumer depth camera in an uncooperative scenario for improved face recognition.
研究成果
The proposed incremental aggregation and refinement method successfully reconstructs higher-resolution 3D face models from low-resolution depth frames, reducing reconstruction error by over 32% compared to single frames and improving face recognition accuracy. It maintains a manageable model size and is suitable for uncooperative scenarios, with potential for extension to other depth cameras.
研究不足
The method assumes small facial expressions; large deformations could negatively affect point position estimation. It is not optimized for real-time processing and may not handle very long-term observations efficiently without further optimization.
1：Experimental Design and Method Selection:
The method involves automatic frame selection based on face pose and distance quality, followed by incremental aggregation and refinement using a Kalman-like estimator to fuse multiple observations while accounting for anisotropic error.
2：Sample Selection and Data Sources:
The Florence 3D Re-Id dataset is used, comprising depth sequences from 16 subjects acquired with a Kinect V2 camera and high-resolution scans from a 3dMD scanner.
3：List of Experimental Equipment and Materials:
Kinect V2 depth camera, 3dMD high-resolution scanner, and a computer for processing.
4：Experimental Procedures and Operational Workflow:
Face detection in RGB frames, extraction of depth data, registration of point clouds using ICP and CPD algorithms, fusion of points based on error covariance, and evaluation of reconstruction accuracy and recognition performance.
5：Data Analysis Methods:
Root Mean Square Error (RMSE) for metric accuracy, Cumulative Matching Characteristic (CMC) curves for recognition accuracy, and statistical analysis of results.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
