2014-04-04 10:30:30 -0600 | commented answer | opencv_createsamples and opencv_traincascade usage example Yeah.. I think that is the problem. There are many tutorials, many (maybe all) have outdated documentation, or work off of datasets that aren't provided. None are "playable". And the tools seem to simply crash when it doesn't like your parameters. So, you quickly hit a brick wall. |
2014-04-04 10:24:09 -0600 | answered a question | How do I run opencv_createsamples with multiple positive images? To my knowledge, there are two ways to use opencv_createsamples (1) Given an info file, listing multiple sample images and the bounding box of the object, it will create the vec file of samples. (2) Given a single image file (using -img), it will generate various distorted versions of that image, and create a vec file of samples. The tool does not allow you to use -img with multiple files, or -info with distortions. One of the guys who created a training tutorial has an (outdated) tool for merging multiple vec files. This is what you are wanting.. unfortunately, i dont know anyone who has the merge tool working these days (there's open questions about that on this site) |
2014-04-04 10:20:17 -0600 | answered a question | Createsamples creating no samples The samples created by opencv_createsamples are collectively stored in the vec file. There's some way to view them, but I haven't done that recently, so I don't remember how. |
2014-04-04 09:59:13 -0600 | commented question | Details of dataset used for training face haarcascades I've ran across a few comments that opencv has not divulged the dataset used for training. One tutorial suggests it's from the FERET database (which I wasn't able to track down). "Probably, the OpenCV developers used the FERET database. It looks that the FERET database became available to download over internet from Jan. 31, 2008(?)." - http://note.sonots.com/SciSoftware/haartraining.html |