Combining Selective Search Segmentation and Random Forest for Image Classification

Random Forest algorithm have been successfully used in many computer vision tasks such as image classification [1] and image segmentation [4]. Recently, Yao et al. showed that a random forest composed of the decision trees where every node is a discriminative classifier outperforms state-of-the-art results in the finegrained image categorization problems [8]. Yao et al. attributed their success to the two main components of their system: discrimination and randomization. Discrimination refers to the use of SVM to learn the splits at each node, whereas randomization refers to a random selection of image patches, which are used as a form of features to learn the splits at each node. There are several problems that may arise from this randomization procedure. Firstly, if we consider image patches of size 50×50 in an 500×500 image, sampling space may contain thousands of patches, which makes it less likely that a randomly selected patch will contain an object of interest for the image categorization. In addition, randomly selected samples are more likely to overlap with each other, which would cause redundancy. Therefore, in this project, I investigated new ways for selecting image patches. In theory, more informative patch selection should result in higher quality splits at each tree node, which in turn should increase overall accuracy of the classifier.