Similar yet visually non-identical objects form recurring patterns that are ubiquitous in the world we live in. Thus an automatic recurring pattern detection algorithm can serve as a stepping stone towards robust higher level machine intelligence. The recognition of such recurring patterns is especially relevant for computer vision since it can lead to saliency detection, image segmentation, image compression and super-resolution, image retrieval and semantically meaningful organization of unlabeled data. This project explores automatic recurring pattern discovery from domain independent images and videos to capture, robustly and flexibly, varying mid-level visual cues emerging from any cluttered background. The work leads to effective and efficient object discovery and scene interpretation. The research team develops an un-supervised method for discovering recurring patterns in a single or multiple images . The key property is the nature of recurring without knowing what recurs. Differing from previous feature- or object-level pairwise-matching-based approaches, recurring pattern discovery from real images is formulated as a joint, 2-dimensional feature assignment optimization problem where multiple objects and multiple feature clusters are considered simultaneously.
The project disseminates the results through publications and sharing data with other researchers. The research of this project contributes to the understanding and capturing of recurring patterns in higher spatial dimensions and spatiotemporal domains. Besides computer vision and computer graphics, many other research fields can also benefit from this research.
