Abstract:
An important visual capability is the ability to quickly detect the presence of multiple similar shapes in an image. We seek a fast method to either determine that no repeated objects are in a image, or to determine that there are repeated objects and identify them. We term this the perceptual clustering property. We segment an image and present a heuristic for the similarity between image segments, and then cluster segments. For segmentation we use the fast MST-based image segmentation presented by Felzenszwalb & Huttenlocher. We define a metric space for segments based on simple per-segment measures including color, size, rough shape etc. Over this space we cluster using a modified hierarchical agglomerative clustering (HAC) algorithm based on shortest distance between clusters. The running time is dominated by the segmentation algorithm, which is itself quite fast; run times are on the order of one second for 240x320 images. We find that our clustering does indeed capture the perceptual clustering property for a wide variety of images. Finally we present future directions for improvement.
|
