Modeling Statistical Relations Among Visual Atoms

Modeling Statistical Relations Among Visual Atoms for Reconstruction Based Encoding

Bag of Words-based methods have achieved promising results in image classification problem. In these methods which are based on a similar algorithm in text categorization, first a set of features are extracted from each image. Then, to treat the problem as a text categorization problem, a set of visual atoms are defined using a subsample of the dataset and each feature is represented as a linear combination of these atoms. These operations are known as dictionary learning and feature coding respectively.

The methods which have been proposed so far are based on an implicit assumption that these atoms are independent of each other. This is while the dependencies among these atoms can help the encoder to code the features more accurately.

 

Figure 1. The effect of dependencies.

 

As a simple example consider the presented figure. In this example, three atoms are learned from the dataset. Assume that the feature being coded has some noise due to feature extraction error, occlusion etc. If an encoder based on reconstructed error is used to code the noisy feature it will use the atoms ‘A’ and ‘B’. This is while if the encoder knows that other similar features (e.g. features which are shown in the bottom of the figure) with high probability use atom ‘C’ when atom ‘A’ is selected,  it can change its decision to increase the consistency.

In this work, we proposed three methods to estimate these dependencies and an efficient technique to incorporate the estimated dependencies in the feature coding stage without changing the original optimization algorithm of the coding methods. The promising results obtained from the proposed method show the importance of these dependencies.

 

People Involved : Mahyar Najibi, Hamid. R. Rabiee, Amirreza Shaban, Ali Soltani-Farani

*The implemented codes will be added to this post.

 

Sparse Signal Processing Group