Classification of the anatomy on 3D scans of human heads | Nicolas Tiaki Otsu
| | Abstract | In order to create full 3D surface models of human heads, a state-of-the art method involves the use of 3D scanner systems, such as the Canfield Scientific Vectra M3 3Dscanner. By utilizing a built-in protocol, partial scans can be formed. The following process of stitching together the partial scans involves a semi-automatic alignment in which salient key points are manually annotated and used for alignment. This process is time consuming and involves a learning curve for the person doing the annotations.
The present work presents a method for automatically assigning anatomical feature labels to the surface vertices of frontal human face scans. Randomized decision forests with weak classifiers have been used as classification models. The models have been trained with a novel method for computing three dimensional vertex feature descriptors, called tangent plane features.
The author has been provided with an active shape model of frontal human faces which is based on scans from 641 test persons from the Danish Blood Donor Study. This has been utilized to generate a large dataset of plausible frontal surface shapes.
The results from the work indicate that the classification of the randomized decision forests is enhanced when the feature computations are based on an area around each surface vertex of up to 10% of the diagonal of the shape bounding box. Setting up a framework for multiple, single-scale investigation has proven to give good insights into the tuning parameters of the randomized decision forests. Cascading classifiers did not improve the results but heuristics for a method that could improve them have been made.
In one of the conducted experiments, the classification of anatomical regions on frontal human face scans by the use of tangent plane features as weak classifiers for training randomized decision forests has yielded an average accuracy of 95% on an independent test set. | | Type | Master's thesis [Academic thesis] | | Year | 2014 | | Publisher | Technical University of Denmark, Department of Applied Mathematics and Computer Science | | Address | Richard Petersens Plads, Building 324, DK-2800 Kgs. Lyngby, Denmark, compute@compute.dtu.dk | | Series | DTU Compute M.Sc.-2014 | | Note | DTU supervisors: associate professor Rasmus Reinhold Paulsen, rapa@dtu.dk, DTU Compute, associate professor Line Katrine Harder Clemmensen, lkhc@dtu.dk, DTU Compute, and PhD student Stine Harder, sthar@dtu.dk, DTU Compute | | Electronic version(s) | [pdf] | | Publication link | http://www.compute.dtu.dk/English.aspx | | BibTeX data | [bibtex] | | IMM Group(s) | Image Analysis & Computer Graphics |
|