| Dave Kauchak
|
4
|
 |
|
05-23-2001 04:20 AM ET (US)
|
|
Boy, that brings me back to the good old days of SATs.
I thought the paper was quite interesting. The explanation of how their system works and the setup is straightforward, but maybe a bit brief. I do think, that they may have constrained their problem a bit given their approach. Although they appear to get good results, I wonder how much better they could do if they tried to generalize the algorithm? Specifically, I would be curious to see if they could gather more information and get better results if instead of a single training pair they used multiple training pairs to learn a given filter. In their model this doesn’t seem appropriate, though.
I also had issues with the experiments that they performed. Although the results seem promising, I have a difficult time actually interpreting what they mean. I understand that this is not the easiest field to actually try and do experiments for because some of the results may be difficult to verify except with the human eye. However, the analysis of the algorithm and the results are still lacking. A helpful thing to provide would be some metric for evaluating the results. This would have allowed them to run on a variety of different images and verify the results. At a minimum, the authors could have had human testers sit down and look at a large sample of results and try and rate how well the filter looked on the B image (even if it was a simple binary answer of good or bad). They then could have presented results from a number of different human verifiers on a decent sized sample.
One thing that I thought the paper did well was advertising the drawbacks and the applicability of the current design. The paper explicitly lays out what they did and what they had accomplished. They also state where they feel their algorithm is lacking and provide a large number of cases where new application and modifications can be tried.
Dave
|
