Last month I attended ISBI 2014 in Beijing as part of the Deconvolution Grand Challenge. These challenges are very important as they encourage cross fertilization of technology.
The top 3 entries were.
Ferreol Soulez – Learn 2D, Apply 3D Method
David Biggs – Olympus Cellsens Deconvolution Software- (Very fast. Dave did his PhD in algorithm acceleration).
Sander Kromwijk – A GPU version of Hessian Schatten Norm Regularization. Also very fast.
I submitted an entry using Total Variation based on Nicolas Dey’s implementation written using the Image2 platform. Theoretically it should been competitive with but slightly below the Hessian Schatten Norm approach (as reported in the Lefkimmiatis’ paper). Though that assumes a near optimal choice for number of iterations and regularization parameter. I used 1000 and 0.0005. 1000 is a lot of iterations, but it is a non-accelerated method and could require even more. The Total Variation method can (and should) be accelerated for example here and here.
In the end they did not reveal the scores for the algorithms (other then the ranking of the top 3). I had mixed feelings about this. On one hand I understand that the organizers did not want to discourage new participants. On the other hand results are often greatly influenced by parameter tuning, so revealing the results could help the participants optimize the parameters as to run their algorithms under the best conditions.
Anyway I would like to thank the organizers Cedrich Vonesch and Stamatios Lefkimmiatis for running the contest. It was a lot of fun to participate in.
Also while in China I had the chance to see a few things. This picture is from a very cool hike I did along the great wall (organized by Beijing Hikers). There was actually a marathon run on the Great Wall while I was in China. The organizers were staying at our hotel, I told them I was a runner and they tried to make me sign up… thankfully I resisted.