David Picard

Short update on JKernelMachines with few news features (new non-convex SVM algorithm, customizable MKL regarding internal SVM solver), many bug fixes, and a more complete example usable as a standalone application.

The reference page can be found on mloss.

We have a paper on image categorization accepted at ICPR next November. This is the other part of the Work Romain Negrel has been doing with VLAT. This time it's about efficiency in image classification. We tried to put every tricks of latest image categorization techniques (like dense sampling, spatial pyramid, and so on) into our VLAT while still retaining small size signatures.

All in all, we managed to achieve 61.5% mAP on VOC2007, which is not bad at all considering we used a single feature and a linear classifier (a stochastic gradient descent from Leon Bottou). Actually, if you put the throttle a bit further, you can expect better results, but then it becomes very heavy computationally speaking. As usual, some code is available here, although it's only for the Holydays dataset right now. At least you can produce the features and then use your own machine learning library (or mine, of course!).

Check the new version: https://mloss.org/software/view/409/

Among the new features worth noting are:

• Improved building procedure
• Evaluation and cross validation packages
• Csv file format
• Lot more of documentation
• Very simple and naive unit testing

I opened an entry on mloss.org for JKernelMachines, as well as a repository on github.

I've also added some sort of documentation in the wiki. I've added a libsvm data format parser and I'm planing a few more new features. Everything is moving a bit fast right now, although the API shouldn't change at all.

Please have a look and make as much comment as you can.

We have a paper accepted at ICIP next september. This is the work Romain Negrel has beeing doing on trying to reduce the size of our VLAT features using some kernel based dimensionality reduction techniques. We focused on search by similarity, and it happened to give very comparable results for large scale benchmarks.

I hope to release some code on this very soon. At least you can check this page

I'll bee in Brugge next week for the ESANN conference.