Abstract
We present an effective image representation based on a new tree-structured coding technique called `random decomposition forests' (RDFs). Our method combines the merits of visual-word based representations and random forests. An RDF is able to decompose a local descriptor into multiple sets of visual words in a recursive and randomized manner. We show that, when combined with standard multiscale and spatial pooling strategies, the code vectors generated by the RDF yield a powerful representation for image categorization, and can achieve state-of-the-art performance on several popular benchmark datasets.