|Ph.D Student||Ilya Blayvas|
|Subject||On Accuracy Analysis of 3D Scanners, Binarization,|
and Machine Learning
|Department||Department of Computer Science||Supervisors||Full Professor Kimmel Ron|
|Full Professor Bruckstein Alfred|
|Full Thesis text|
The first part of our research deals with the accuracy analysis of the optical 3D scanners, while the second part describes the related algorithms, that were developed during the research: binarization and machine learning.
We start from the physical analysis of depth from focus/defocus and depth from structured light systems, which are popular and efficient active optical 3D scanners. The 3D scanning system accuracy is defined by the hardware, object properties and the processing algorithm. The system analysis had shown, that the system optics has a crucial role on the accuracy, while the role of processing algorithm is less significant.
The derived accuracy estimators allow to predict an accuracy of the given system, or to design a new system, complying with the pre-defined accuracy requirements.
Unfortunately some of the important publications in the field failed to report crucial optical parameters, and therefore the described algorithms can not be compared in the optics-invariant way. Therefore, our result can be considered not only as a contribution to the field of optical 3D acquisition, but also as a message to the computer vision community, that the role of optics is crucial, and the optical systems should be analyzed or at least completely described in order to allow the hardware-independent algorithm comparison between the different publications.
We continue with the physical analysis of the structured light 3D scanner, where we discuss both the role of optics and the image sensor on the accuracy. Structured light 3D scanners with binary patterns is one of the efficient and popular 3D scanning approaches. Image Binarization is a crucial processing step in these scanners, when the black and white regions of the pattern are determined from the acquired non-uniform gray level image.
We had re-visited the image binarization problem, and proposed our analysis and improvement to the popular Yanovitz-Bruckstein binarization method.
Our binarization algorithm is based on efficient construction of the adaptive threshold surface for image binarization.
The threshold surface is constructed as a multi-resolution interpolation of sparse support points.
The generalization of the construction algorithm to multi-dimensional space yielded a multi-resolution density estimator, that allowed to construct a computationally efficient Bayesian classifier, with constant time per training point and per query.