1 | initial version |
Yes it is possible both options: - 1 xml for 1 category - 1 xml for many categories.
The rigid objects approach have some limitations (ie: you cannot detect many cateogories with wide variability). Anyway it is possible to do if you consider the object in constrained scenarios. Both of approaches have pros and cons.
The support for traincascade.exe is not guaranteed with the latest versions of OpenCV. The open source world is moving massively to deep learning approaches without considering the "old school" methods are (still in 2020) the best trade-off in speed/accuracy for a lot of application. Anyway you can use earlier versions of OpenCV to train cascades, I'm still using the traincascade in Opencv 2.4.10 and it is still perfectly working and suitable for OpenCV 3 and 4. More here:
Regards!