Logo image
Back
A convolutional neural network-based approach to personalized 3D modeling of the human body and its classification
Thesis   Open access

A convolutional neural network-based approach to personalized 3D modeling of the human body and its classification

Semanti Basu
Master of Science (M.S.), Drexel University
Jun 2020
DOI:
https://doi.org/10.17918/00000157
pdf
Basu_Semanti_20202.76 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Three-dimensional display systems Human body--Mathematical models Neural networks (Computer science) Three-dimensional imaging Pattern recognition systems
In this thesis, we introduce an integrated method to build personalized full body 3D models of people given frontal and profile silhouette images. Several deep convolutional neural network (CNN) architectures have been designed and trained to accurately estimate the positions of a set of anthropometric set of ordered control points on the frontal and profile silhouette images. For the prediction of key points on the frontal silhouette image, the output from four different convolutional neural networks have been fused together to generate the final coordinates. A global CNN is first designed to predict those control points on all parts of the body. This has been reinforced with local deep CNN architectures focused on the prediction of control points on localized areas of the body to improve on the accuracy of predictions. Fusing the global and local predictions yielded an estimate of the coordinates of 56 control points on the frontal image and 26 control points on the side view image of a person. The controlled points are then regularized to reside on the silhouette of the frontal and profile images using a combination of Canny edge detector and shortest distance mapping. The set of regularized control points are then fed into a model-based 3D reconstruction algorithm [1] to yield the corresponding high-resolution 3D model of the person. A database of 800 models from the Caesar dataset were studied, of which 100 were used to train and the other 700 were used for testing and classification of 3D models. Our method achieves an accuracy of 99.7 % in prediction of control points and 3D reconstruction using those points. We also present a classification scheme to allocate a test surface to one of competing base surfaces. The classification is based on computing the error between salient points with identical anthropometric meaning that reside on a nested set of boundaries in the frontal and profile projection image spaces. The method can have a variety of applications ranging from medical imaging, to 3D modeling for recognition, virtual reality, generation of video games, 3D animation, etc.

Metrics

69 File views/ downloads
23 Record Views

Details

Logo image