T. Pfrommer, "Improving 3D Registration Results of Foot Models Dramatically with a Machine Learning Enhanced Geometric Feature Extraction", Proc. of 3DBODY.TECH 2022 - 13th Int. Conf. and Exh. on 3D Body Scanning and Processing Technologies, Lugano, Switzerland, 25-26 Oct. 2022, #43, https://doi.org/10.15221/22.43.

Title:

Improving 3D Registration Results of Foot Models Dramatically with a Machine Learning Enhanced Geometric Feature Extraction

Authors:

Tobias PFROMMER

ShoeFitter GmbH, Konstanz, Germany

Abstract:

In this paper, a method is presented that enables a true-to-scale reconstruction of 3D foot models using the iPhone's Face ID sensor. For this purpose, multiple incoming 3D point clouds representing the foot are registered piece by piece with each other. A feature-based registration pipeline is used for pairwise registration. Geometric feature extraction in such pipelines is the first and most important step for correct registration of two 3D point clouds. For this purpose, we train and apply learned feature descriptors based on Fully Convolutional Geometric Features (FCGF). It is shown that the features computed by our trained feature extractor are more robust and faster than conventional methods. We trained FCGF using a self-generated dataset of 3D foot models augmented with synthetic data. The trained feature model was optimized with hyperparameters. For better visualization of the high-dimensional features, a t-SNE-based visualization is used to assign features that are reliably found in the same location of the foot in different models. Based on the detected features, the optimal transformation of two point clouds is estimated by a feature-based RANSAC algorithm. In the benchmarks, it is found that the implemented feature descriptor consistently achieves better feature matching and registration recall results than comparable feature descriptors. With the final trained model of the feature descriptor within the presented registration pipeline, a 3D reconstruction of a foot can be performed using an overlap of only 27 percent. This makes the reconstruction of the 3D model much more robust than using comparable state-of-the-art methods.

Keywords:

3d body scanning, point cloud registration, foot measurement, feature descriptor, 3d reconstruction, deep learning, fully convolutional network, 3d foot model

Details:

Full paper: 2243pfrommer.pdf
Proceedings: 3DBODY.TECH 2022, 25-26 Oct. 2022, Lugano, Switzerland
Paper id#: 43
DOI: 10.15221/22.43
Presentation video: 3DBodyTech2022_43_pfrommer.mp4

Copyright notice

© Hometrica Consulting - Dr. Nicola D'Apuzzo, Switzerland, hometrica.ch.
Reproduction of the proceedings or any parts thereof (excluding short quotations for the use in the preparation of reviews and technical and scientific papers) may be made only after obtaining the specific approval of the publisher. The papers appearing in the proceedings reflect the author's opinions. Their inclusion in these publications does not necessary constitute endorsement by the editor or by the publisher. Authors retain all rights to individual papers.

Note: click the + on the top left of the page to open/close the menu.