Benjamin Kimia (Brown University)
The bottom-up “segmentation followed by recognition” strategy has for some time now given way to feature-based discriminative recognition with significant success. As the number of categories and exemplars per category increases, however, low-level features are no longer sufficiently discriminative, motivating the construction and use of more complex features. It is argued here that these complex features will necessarily be encoding shape and this in turn requires curves and regions, thus reviving aspects of bottom-up segmentation strategies. We suggest that the demise of segmentation was due to prematurely committing to a grouping in face of ambiguities and propose a framework
for representing multiple grouping options in a containment graph. Specifically, we use contour symmetry to partition the image into atomic fragments and define transforms to iteratively grow these atomic fragments into mote distinctive perceptual fragments, the nodes of the containment graph. We also briefly present a fragment-based language for generating shapes and the use of fragments in top-down category recognition. The bottom-up and top-down processes are then integrated by interaction through the mid-level representation of perceptual fragments.