I don't understand why the computer vision community is so locked in to matlab. Sure it's a great tool, but it's only marginally better than what the python community has to offer (that margin being the network effect I'm about to rage against) and it's damn expensive. You'd think the high cost and vendor lock-in would be off-putting to academics (the discounts are such a transparent ploy.) I mean, what good is it releasing some code under the GPL when you need a $2000 system (plus a handful of toolboxes at $1000 a pop, plus maintenance fees) to run it.

What's particularly galling (to a matlab hater like me, anyway) is how some of the most highly-cited and useful work that's made available by researchers is done so only through matlab. The de facto standard benchmark for keypoints (http://www.robots.ox.ac.uk/~vgg/research/affine/) is a matlab function, but it's not like that code uses matlab in any interesting way, it just reads text files. And even though it's not computationally intesive (the benchmark is only on small set of images) there is a mex function involved to do the actual computation, which makes it a pain in the ass to use with octave. Basically they're using matlab as a $2000 bash shell. Same situation for the standard segmentation benchmark (http://www.eecs.berkeley.edu/Research/Projects/CS/vision/bsd...) And then there's LabelMe (http://labelme.csail.mit.edu/) the "open" annotation tool. They've got a really slick javascript UI for crowdsourcing ground truth collection for image segmentation / object classification tasks. They make the data freely available (which is awesome) but to access it you have to go through a matlab toolbox...? WTF?! I mean really, you've obviously got someone on your team who knows a thing or two about the web: take a page out of Flickr's playbook (after all they've done for the CV community) and make your data available via open standards. There's not even a goddamned download link on the page. It's like they're on the Mathworks payroll.

Grrr...