Supplementary material:

- 2moons.gif: GIF animation, for the dataset of noisy 2-moons experiment
  of the paper. In this experiment, we run the homotopy version of our
  Laplacian K-modes algorithm and decrease kernel width from 5 to 0.1 in
  10 steps geometrically while fixing the trade-off parameter lambda=1.
  We show in the left panel the clustering result and kernel density
  estimate at sigma changes, and in the right panel the out-of-sample
  assignment mapping.

- occluder.gif: GIF animation, for the occluder segmentation experiment of
  the paper. In this experiment, we fix the graph (each pixel is connected
  to its surrounding 8 pixels), and compare the segmentation result of
  normalized cut (left panel) and Laplacian K-modes (right two panels) at
  different Gaussian kernel widths using K=5. The goal is to have the
  occluder (denoted by the black box) segmented out as a separate cluster. 
  The soft assignments plot the assignment (as a greyscale from 0 to 1)
  corresponding to the cluster most associated with the occluder.

All these animations may be seen with a web browser or with specialized
GIF image viewers.