Hi, On Mon, Aug 29, 2011 at 12:27 PM, Emmanuelle Gouillart <emmanuelle.gouillart@normalesup.org> wrote:
having opencv as a dependency has two important drawbacks for me. First, it makes the installation of the scikit more difficult.
There is certainly a cost; the question is whether that cost comes with enough advantage to make it worth our while. There are currently binary packages available on Windows, in most of the major distros (and on OSX building is supported, although I saw no binary package).
Second, the documentation of openCV does not match the standards of numpy/scipy docstrings (no examples, few references to articles).
When we use it as a dependency, I don't think this needs to matter too much. We'll be using it to do things like fast convolution, feature finding, colour conversions, etc. We'll still have to write documentation for anything we want to expose to our users.
Performance is great, but I think we'll gain more users by putting effort on user-friendliness than if we put it on performance. For examples, a large fraction of my colleagues are using image processing, but on Windows, and they get discouraged by a package with unsufficient documentation.
We have two groups of users, and their needs are somewhat different. Personally, I need accurate algorithms for research, but I know that many others need speed for doing near real-time processing, for example. Regards Stéfan