Figure 1

Region partitioning of toy image data. (a) Original RGB image to which Gaussian noise was added to create (b) noisy RGB image. Image visualization of classes obtained by replicated k-means clustering [15] of all the pixels via (c) original noisy RGB space, and (d) graph embedding [6] of noisy RGB data. 2D plots of (e) R-G, (f) R-B, and (g)G-B planes are also shown where colors of objects plotted correspond to the region in (b) that they are derived from. The discriminatory 2D spaces ((e) and (f)) are combined via consensus embedding, and the visualized classification result is shown in (h). Note the significantly better image partitioning into foreground and background of (h) compared to (c) and (d).