Decoupled BoF Histograms: using substantial homogeneity

BACK
In histogram-based methods, such as BoF, content of image is represented by distribution of visual words over the whole image. This is not necessarily efficient to assign all the visual words to a single histogram. In order to improve discriminancy of BoF histograms, we try to decouple the histograms and distribute visual words into several more discriminant histograms in a smart way.

We use a texture-transform method to find the regions with similar substantial properties. Then BoF histograms are built for each region independently. We can study response of each region to different types of feature descriptors. Besides, it is interesting to see contribution of each region in image classification performance when histograms from all of the regions are combined (concatenated) together.
Figure 1 illustrates how the images are segmented into different regions. It is interesting to see that how the regions with the same color are similar in terms of their roughness.

We use Multiple Kernel Learning method to combine and also analyse contribution of different feature descriptors in different image regions.
Figure 1. First row: Original images, second row: Response of texture-transform, third row: Decomposition of images into three regions by thresholding response of the texture-transform.
Related publications: