The PR #1223 created a workaround to ignore the max feature limit for PyDetector. The proper fix would be to filter the already generated descriptors from PyDetector at the same time we limit the keypoints.