Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Soft biometrics, biomedical sciences, and many other fields of study pay particular attention to the study of the geometric description of the human body, and its variations. Although multiple contributions, the interest is particularly high given the non-rigid nature of the human body, capable of assuming different poses, and numerous shapes due to variable body composition. Unfortunately, a well-known costly requirement in data-driven machine learning, and particularly in the human-based analysis, is the availability of data, in the form of geometric information (body measurements) with related vision information (natural images, 3D mesh, etc.). We introduce a computer graphics framework able to generate thousands of synthetic human body meshes, representing a population of individuals with stratified information: gender, Body Fat Percentage (BFP), anthropometric measurements, and pose. This contribution permits an extensive analysis of different bodies in different poses, avoiding the demanding, and expensive acquisition process. We design a virtual environment able to take advantage of the generated bodies, to infer the body surface area (BSA) from a single view. The framework permits to simulate the acquisition process of newly introduced RGB-D devices disentangling different noise components (sensor noise, optical distortion, body part occlusions). Common geometric descriptors in soft biometric, as well as in biomedical sciences, are based on body measurements. Unfortunately, as we prove, these descriptors are not pose invariant, constraining the usability in controlled scenarios. We introduce a differential geometry approach assuming body pose variations as isometric transformations of the body surface, and body composition changes covariant to the body surface area. This setting permits the use of the Laplace-Beltrami operator on the 2D body manifold, describing the body with a compact, efficient, and pose invariant representation. We design a neural network architecture able to infer important body semantics from spectral descriptors, closing the gap between abstract spectral features, and traditional measurement-based indices. Studying the manifold of body shapes, we propose an innovative generative adversarial model able to learn the body shapes. The method permits to generate new bodies with unseen geometries as a walk on the latent space, constituting a significant advantage over traditional generative methods.
Piccirilli, Marco, "Machine Learning Approaches to Human Body Shape Analysis" (2018). Graduate Theses, Dissertations, and Problem Reports. 6417.