Probabilistic Compositional Active Basis Models for Robust Pattern Recognition

Adam Kortylewski and Thomas Vetter

Abstract

Hierarchical compositional models (HCMs) have shown impressive generalisation capabilities, especially compared to the small amounts of data needed for training. However, regarding occlusion and other non-linear pattern distortions, experimental setups have been controlled so far. In this work, we study the robustness of HCMs under such more challenging pattern recognition conditions. Our contribution is three-fold: First, we introduce a greedy EM-type algorithm to automatically infer the structure of compositional active basis models (CABMs). Second, we formulate the proposed representation and its learning process in a fully probabilistic manner. Finally, building on the statistical framework, we augment the CABM with an implicit geometric background model that reduces the models sensitivity to pattern occlusions and background clutter. In order to demonstrate the robustness of the proposed object representation, we evaluate it on a complex forensic image analysis task. We demonstrate that probabilistic CABMs are capable of recognising patterns under complex non-linear distortions that can hardly be represented by a finite set of training data. Experimental results show that the forensic image analysis task is processed with unprecedented quality.

Session

Posters 1

Files

PDF iconExtended Abstract (PDF, 1M)
PDF iconPaper (PDF, 2M)

DOI

10.5244/C.30.30
https://dx.doi.org/10.5244/C.30.30

Citation

Adam Kortylewski and Thomas Vetter. Probabilistic Compositional Active Basis Models for Robust Pattern Recognition. In Richard C. Wilson, Edwin R. Hancock and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 30.1-30.12. BMVA Press, September 2016.

Bibtex

        @inproceedings{BMVC2016_30,
        	title={Probabilistic Compositional Active Basis Models for Robust Pattern Recognition},
        	author={Adam Kortylewski and Thomas Vetter},
        	year={2016},
        	month={September},
        	pages={30.1-30.12},
        	articleno={30},
        	numpages={12},
        	booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
        	publisher={BMVA Press},
        	editor={Richard C. Wilson, Edwin R. Hancock and William A. P. Smith},
        	doi={10.5244/C.30.30},
        	isbn={1-901725-59-6},
        	url={https://dx.doi.org/10.5244/C.30.30}
        }