Reconstruction and Recognition of 3D Heads Employing Hybrid Principal Component Analysis

Q. Wu and J. Ben-Arie (USA)


Machine Learning, Signal and Image Processing, Hybrid Principal Component Analysis (HPCA), 3D head recon struction and recognition


This paper proposes an extension to the method of recogni tion by appearance of faces and objects based on the Prin cipal Component Analysis (PCA). While the methods of recognition by appearance are two dimensional, we pro pose an extension to three dimensional (3D) recognition by transforming the 2D image data into 3D image. This novel transformation is performed using a Hybrid Principal Com ponent Analysis (HPCA) i.e. hybrid Hotelling - KL trans form. In our approach, we reconstruct 3D range images from 2D gray scale images by HPCA. HPCA can be con sidered as a learning algorithm where the system learns to relate 3D head structures to their 2D appearances (i.e. their images). In the learning stage, the algorithm is trained by a set of head images with corresponding 3D data. In the recovery stage, the algorithm is presented with new 2D im ages and yields 3D head structure reconstructions. At the recognition phase, the similarity of the reconstructed 3D head to each one of the 3D models is measured. The simi larity is inversely proportional to the average matching er ror. We find the average error using the Iterative Closest Point (ICP) algorithm, which first finds the optimal geo metric head-model transformation parameters (translation, rotation and scale). Next, it computes the average match ing error by averaging the minimal distances of every 3D head point to the transformed model. Experimental results demonstrate that 2D images are accurately reconstructed and provide high recognition rates.

Important Links:

Go Back