From Noisy Point Clouds to Complete Ear Shapes: Unsupervised Pipeline

Abstract

Ears are a particularly difficult region of the human face to model, not only due to the nonrigid deformations existing between shapes but also to the challenges in processing the retrieved data. The first step towards obtaining a good model is to have complete scans in correspondence, but these usually present a higher amount of occlusions, noise and outliers when compared to most face regions, thus requiring a specific procedure. Therefore, we propose a complete pipeline taking as input unordered 3D point clouds with the aforementioned problems, and producing as output a dataset in correspondence, with completion of the missing data. We provide a comparison of several state-of-the-art registration and shape completion methods, concluding on the best choice for each of the steps.

Probabilistic Registration for Gaussian Process 3D shape modelling in the presence of extensive missing data

“…Furthermore, the dataset and template were subsampled with respect to the original datasets, in order to facilitate computation. We also consider here the fitting of a subset of shapes, as the dataset is composed from 500 samples, but it does not present much variability across them (see our previous work [23] for a more detailed explanation.) Therefore, we consider the shapes with more deformation with respect to the template (measured as the average of Euclidean distance between corresponding shape points).…”

“…Our approach relates to this method, in the formulation of the problem but not in the way the correspondences are retrieved. For ear shapes, given the large regions of missing data and the highly non-rigid deformations, the closest point approach often leads to undesirable results [23].…”

“…The former is a representative example of the shape being studied, as close as possible to any other shape, without missing data, outliers or noise, while the latter is any sample of the shape we have obtained. In previous work [23] with ear data scans, we have shown that the most part of the effects of translations and rotations can be previously removed; so it is here considered that the target and template were already preprocessed. Additionally, we want to introduce the scaling in the model, as this represents differences in size of the ears.…”

“…how to bring together models originated from different datasets. When applying this process to 3D ear point clouds, we have observed that indeed this is the most promising approach [23], but it calls for an improved outlook on the registration procedure. When models of the full head are considered, ears are a small detail that is often overlooked or disregarded [2].…”