Best approach to detect different known objects

because both cascadés and hog descriptors will degrade

definitly not (and same problem for cascadés and hog descriptors)

great to get a homography between scenes,
terrible at discriminating object classes

great for CBIR / knn search, but (probably) not what you wanted.


please do some more research on cnns, and have a look at [the android detection sample](https://github.com/opencv/opencv/tree/master/samples/android/mobilenet-objdetect)