NIPS 2004
Spectral clustering has been theoretically analyzed and empirically proven useful. There are still open issues:
(i) Selecting the appropriate scale of analysis,
(ii) Handling multi-scale data,
(iii) Clustering with irregular background clutter, and,
(iv) Finding automatically the number of groups.
We explore and address all the above issues. We first propose that a `local’ scale should be used to compute the affinity
between each pair of points. This local scaling leads to better clustering especially when the data includes multiple scales and when the clusters are placed within a cluttered background. We further suggest exploiting the structure of the eigenvectors to infer automatically the number of groups. This leads to a new algorithm in which the final randomly initialized k-means stage is eliminated.Paper in pdf
Matlab code and data for synthetic experiments
Experiments:
Image segmentation results, based on gray-scale differences alone. The number of clusters was set manually here to force a large number of clusters. Since the scale is tuned locally for each pixel we obtained segments with both high and low contrast to the surrounding. We compare our result with those obtained using a single scale (set manually) and k-means (algorithm of Jg-Jordan-Weiss). To have a fair comparison we chose a scale which worked well for one of the images and used the same scale for all other images. Both algorithms were asked to provide the same number of clusters. Using a single scale we usually got a smaller number of clusters than asked for as due to low sensitivity some clusters were determined by the final k-means to be redundant.
| Our result (self-tuning) | Single Scale Segmentation |
For this image we got good results using both approaches. |
 |
Recall, we use intensity only and no texture information. |
Using the same scale as above failed on this image. |
| Using local-scaling the high contrast trees as well as the low contrast windows and building boundaries are detected.  |
Only high contrast boundaries are detected.
|
 |
 |
| Both low contrast and high contrast boundaries are detected. The sky is over segmented.  |
The lower contrast round roof boundary was missed.
|
Experiments on synthetic data
These are fully automatic results: group number and scale recovered automatically!!
