Hybrid Fusion: Beyond Early and Late Fusion for Texture Classification
New Trends and Challenges in Computer Vision: Progress of Research and Development - 2009
This paper presents a novel method for combining
multiple attributes in order to classify the different
categories. We start by providing a detail analysis
of how to optimally fuse color and shape information
for texture classification. For this reason we
analyze the two existing approaches, called early
and late fusion, and argue that both approaches
are suboptimal for some classes. To overcome
this shortcoming, we propose to merge the two
approaches into a single combined early and late
fusion representation of an image. We further
propose to combine this new hybrid fusion with
a texture representation in an efficient way. Experiments
have been conducted on a large dataset
of ten different image categories and the results
show that all these three cues are important for
the task of texture classification and our proposed
method increases the overall performance significantly.
Images and movies
BibTex references
@InProceedings\{SVV2009a, author = "Fahad Shahbaz Khan and Joost van de Weijer and Maria Vanrell", title = "Hybrid Fusion: Beyond Early and Late Fusion for Texture Classification", booktitle = "New Trends and Challenges in Computer Vision: Progress of Research and Development", year = "2009", abstract = "This paper presents a novel method for combining multiple attributes in order to classify the different categories. We start by providing a detail analysis of how to optimally fuse color and shape information for texture classification. For this reason we analyze the two existing approaches, called early and late fusion, and argue that both approaches are suboptimal for some classes. To overcome this shortcoming, we propose to merge the two approaches into a single combined early and late fusion representation of an image. We further propose to combine this new hybrid fusion with a texture representation in an efficient way. Experiments have been conducted on a large dataset of ten different image categories and the results show that all these three cues are important for the task of texture classification and our proposed method increases the overall performance significantly.", url = "http://www.cat.uab.cat/Public/Publications/2009/SVV2009a" }