Anthropometric salient points and convolutional neural network (CNN) for 3D human body classification

Semanti Basu, Chenxi Li, Fernand S. Cohen

Abstract: In this paper, we introduce a 3D body shape biometric for recognizing a person as one of C possible individuals stored in a database in 3D free form (scatter of 3D data points) using a couple of canonical images (front and side) taken of that individual. The first step is to reconstruct the full body 3D shape model of the individual based on their frontal and profile silhouette images. This is done by using a 3D generic model that gets morphed in accordance with the canonical images of the individual. Starting with a small set of anthropometric interconnected ordered intrinsic control points residing on the silhouette boundary of the projections of the generic model onto the frontal and profile image spaces, corresponding control points on two canonical images of the person are automatically found. This imports equivalent saliency between the two sets. The positions of these control points on the canonical images are attained using deep convolutional neural networks (CNNs) that have been trained offline on a set of images of different individuals. Further equivalent saliencies between the projected points from the generic model and the canonical images are established through loop-subdivision. To personalize the generic model, points on the generic model are morphed to be consistent with their equivalent points on the canonical images. The 3D reconstruction yields sub-resolution errors when tested on the CAESAR data set with 700 different individuals. Classification based on the error between salient points with identical anthropometric meaning residing on nested sets of boundaries in the frontal and side projections, achieves an accuracy of 96%. This is to be compared to an accuracy of 72% when using KNN nearest distance point classification between test and base models.