Abstract
More often than not, visual data objects, such as images, can be described by multiple
features due to its multi-view nature, with each view corresponding to an individual
feature type. For example, an image can be described by a color view and a shape view.
Commonly, these views provide complementary information to each other, which can
lead to superior performance over single view based models by combining heterogeneous
views. One critical challenge of successfully leveraging the complementary information lies
in how to explore the consensus among all views. To address such challenge, we propose
a series of novel methods to achieve the multi-view consensus so that the complementary
information from multiple views can be seamlessly leveraged. Specifically, we propose to
leverage pair-wise similarity metric as well as high order similarity from multiple views
for visual objects retrieval, semi-supervised based manifold ranking and multi-view based
hashing similarity search over large scale visual data objects. To show the significance of
our techniques, we further apply them to applications such as clustering, mode seeking
and salience detection. We experimentally validate the effectiveness of our proposed multi-view based
methods, which can well achieve the multi-view consensus on both synthetic and real-world
datasets. It demonstrates that our proposed methods outperform both existing multi-view and single view based methods.