Context-aware discovery of visual co-occurrence patterns

Once images are decomposed into a number of visual primitives, it is of great interests to cluster these primitives into mid-level visual patterns. However, conventional clustering of visual primitives, e.g., bag-of-words, usually ignores the spatial context and multi-feature information among the visual primitives and thus cannot discover mid-level visual patterns of complex structure. To overcome this problem, we propose to consider both spatial and feature contexts among visual primitives for visual pattern discovery in this chapter. We formulate the pattern discovery task as a multi-context-aware clustering problem and propose a self-learning procedure to iteratively refine the result until it converges. By discovering both spatial co-occurrence patterns among visual primitives and feature co-occurrence patterns among different types of features, the proposed method can better address the ambiguities of visual primitives.