Non-rigid 3D Shape Matching using Outlier-robust Spectral Embedding (Niet-rigide 3D-vormovereenstemming door uitschieter-robuuste spectrale inbedding)
Non-rigid 3D Shape Matching using Outlier-robust Spectral Embedding
Smeets, Dirk; S0109145
Translating physical and biological processes into computer models allows to simulate and analyse the real world and, as such, obtain a better understanding. This thesis addresses the comparison of 3D shapes based on visual data, especially in the presence of intra-subject deformations including pose variations, and focusses on the applications of 3D face recognition and 3D object retrieval.The proposed approach consists of the embedding of each shape into a high-dimensional canonical domain resulting in an intra-subject deformation invariant embedding shape, which encodes the inter-point distances of the original shapes according to a predefined metric. The choice of the metric determines the extent of invariance for intra-subject deformations. By comparing these embedding shapes, a global shape similarity is assessed and point correspondences are obtained. Robustness against outliers and missing data is achieved by a weighted spectral embedding, suppressing outliers which are estimated by a multi-scale local feature method. As a result, salient surface features are automatically detected, described and matched.With the methods developed, excellent results were obtained on reference data for expression-invariant face recognition as well as on international challenges on non-rigid 3D object retrieval.From this research, we conclude that the developed spectral embedding techniques offer an important contribution to fully automated comparison of 3D shapes, in presence of intra-subject variations and outliers.
Smeets D., ''Non-rigid 3D shape matching using outlier-robust spectral embedding'', Proefschrift voorgedragen tot het behalen van het doctoraat in de ingenieurswetenschappen, K.U.Leuven, May 2012, Leuven, Belgium.
Table of Contents:
List of symbols
1. General introduction
2. Intra-subject deformation modelling: state-of-the-art
3. General approach
5. Spectral embedding
6. Higher-order spectral embedding
7. Outlier robust spectral embedding
8. Conclusions and future work