| Who | When |
Messages | |
|
|
|
|
|
| sitctaget
|
11
|
 |
|
08-16-2008 08:49 PM ET (US)
|
|
dargetcaal | | |
|
| wowgold
|
10
|
|
|
08-03-2008 04:05 AM ET (US)
|
|
|
| jackbon
|
9
|
|
|
07-08-2008 01:59 AM ET (US)
|
|
|
| JZY
|
8
|
|
|
07-01-2008 10:47 PM ET (US)
|
|
|
|
|
7
|
|
|
05-16-2008 12:34 PM ET (US)
|
|
Deleted by topic administrator 05-17-2008 10:16 AM
|
| Taylor
|
6
|
|
|
07-21-2006 04:51 PM ET (US)
|
|
Great discussion, lotsa funny links and stuff. I enjoyed my stay and will come back again. You can visit my discussions buy pain killers in uk webpage devoted to buy pain killers in uk. lasix plus webpage devoted to lasix plus. too. |
| Tomasz Malisiewicz
|
5
|
|
|
02-27-2006 11:09 AM ET (US)
|
|
I'd like to make one comment on the idea of obtaining multiple segmentations by grouping superpixels via a pairwise affinity function. We already saw this approach in the "Geometric Context from a Single Image" paper.
This idea of utilizing a pairwise affinity function and a base superpixel segmentation is powerful and was used in the "Graph Partition by Swendsen-Wang Cuts" 2003 paper to speedup their Markov Chain Monte Carlo procedure. In the Swendsen-Wang Cuts paper, the idea was to use medium-sized regions as candidates for splitting and merging as opposed to the small superpixel-sized regions used in the DDMCMC paper. By using these medium-sized regions they were able to obtain a 20-40 time speedup over the original DDMCMC approach.
The approach of Hoiem et al is intuitive and easy to implement while the MCMC papers are overly mathematical and not straightforward to implement. I think some mathematical formalism is necessary when trying to incorporate multiple segmentations, but perhaps one does not really need to sample the space of all segmentations of an image (by abandonding boundary diffusion) and use all of heavy machinery of the DDMCMC paradigm.
|
| Tomasz Malisiewicz
|
4
|
|
|
02-27-2006 10:30 AM ET (US)
|
|
When reading this paper, a key thing to remember (mentioned in the introduction) is that the term "particle" refers to weighted samples which encode nonparametric probabilities (parzen-window style) in various subspaces.
|
| Tomasz Malisiewicz
|
3
|
|
|
02-22-2006 07:26 PM ET (US)
|
|
|
| Tomasz Malisiewicz
|
2
|
|
|
02-19-2006 05:49 PM ET (US)
|
|
|
Dave Bradley
|
1
|
|
|
02-05-2006 05:34 PM ET (US)
|
|
Pleas post you thoughts on "Image Segmentation by Data-Driven Markov Chain Monte Carlo" here.
|
