Learning Additive Kernel For Feature Transformation and Its Application to CNN Features

Takumi Kobayashi

Abstract

Feature transformation is an important process following feature extraction for improving classification performance. It has been frequently addressed in the kernel-based framework utilizing non-linear kernel functions, and the additive kernel equipped with explicit feature mapping works as efficient and effective (non-linear) feature transformation. The kernel functions, however, are defined in a top-down manner taking into account the inherent nature of the features, which makes it difficult to appropriately apply them to the features whose characteristics are not fully disclosed, such as CNN features. In this paper, we propose a method to learn an additive kernel of which explicit mapping serves feature transformation. By means of a bottom-up learning approach leveraging annotated data, the proposed method builds the kernel function of high generality and discriminative power even for the CNN features. The experiments on various datasets using various types of pre-trained CNN features show favorable performance improvement by the learned additive kernel (feature transformation) of which generality over the datasets and the CNN models is also demonstrated.

Session

Posters 2

Files

PDF iconExtended Abstract (PDF, 378K)
PDF iconPaper (PDF, 1M)

DOI

10.5244/C.30.98
https://dx.doi.org/10.5244/C.30.98

Citation

Takumi Kobayashi. Learning Additive Kernel For Feature Transformation and Its Application to CNN Features. In Richard C. Wilson, Edwin R. Hancock and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 98.1-98.12. BMVA Press, September 2016.

Bibtex

        @inproceedings{BMVC2016_98,
        	title={Learning Additive Kernel For Feature Transformation and Its Application to CNN Features},
        	author={Takumi Kobayashi},
        	year={2016},
        	month={September},
        	pages={98.1-98.12},
        	articleno={98},
        	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.30.98},
        	isbn={1-901725-59-6},
        	url={https://dx.doi.org/10.5244/C.30.98}
        }