Grouping-based Hypothesis Verification in Object Recognition

e. Few wrong f. Hypothesis No. 32. g. Additive score h. Grouping-based score hypotheses (the correct one) for hypotheses 1 ? 45. for hypotheses 1 ? 45. Figure 3: Example 2: The input is a real image (a), and its edges are shown in (b). The object is an apple, its boundary modeled by a simple hand-drawn 2D spline. The grouping information, represented by the set of tested grouping cues (c), and their results (d), emphasize the smooth edges. Few of the wrong hypotheses, (e), and the correct one, (f), are presented by their associated Object set of edge points, superimposed on the original image. 45 hypotheses were veriied by an additive score (g) and by the grouping based score (h). The additive scores of 31 wrong hypotheses, numbered 1 ? 31, are higher than the score of the correct one (the 32nd hypothesis). With the grouping-based score, the correct hypothesis got the highest score, which is clearly much higher than all others. In this real situation, the erroneous hypotheses are created by a random accumulation of clutter in some regions of the edge image, which is a mixture of the the two failure mechanisms presented in Figure 1(c,d). Still the incorrect hypothesis are rejected, mainly by the negative contribution of the missed cues between edge pixels of the hypothesized boundary.