Efficient Object Localization with Variation-Normalized Gaussianized Vectors - Intelligent Video Event Analysis and Understanding - page 108

Fig. 4 Examples of good and bad localization based on Gaussianized vector representation.
(The black and white bounding boxes in the images are the ground truth and the hypotheses
respectively. Best viewed in color.)

6Con lu ion

In this chapter, we discuss effective object localization leveraging efficient and ef-
fective searching method, and robust image representation and learning method. In
particular, we present an efficient object localization approach based on the Gaus-
sianized vector representation. We design a quality bound for rectangle sets char-
acterized by the Gaussianized vector representation. This bound can be obtained
for any rectangle set within the image boundaries, with little extra computational
cost, in addition to calculating the Gaussianized vector representation for the whole
image classification problem. Adopting the branch-and-bound search scheme, we
leverage the proposed quality bound for fast hierarchical search.
We further incorporate a normalization approach that suppresses the within-class
variation, by de-emphasizing the undesirable subspace in the Gaussianized vector
representation kernels. This helps achieve improved robustness to variation in the
object class and the background.
The proposed object localization approach based on the Gaussianized vector rep-
resentation outperforms a similar localization system based on the generic

histogram-of-keywords representation on a multi-scale car dataset.

Acknowledgements. This research is funded by NSF grants IIS 08-03219 and IIS-0703624.