While working with CLandmarksMap using stereo vision data in an RBPF slam framework, I found that it runs too slow and it slows down considerably when increasing the number of particles. Investigation showed that for a single observation, CLandmarksMap::loadSiftFeaturesFromStereoImageObservation() is called 2 times for each particle, which is a time consuming operation. I implemented a simple hack by caching a map built from an observation and used that map instead of calling the mentioned function when it was needed for the same observation. I think, some kind of caching mechanism should be implemented in CLandmarksMap (or maybe a more suitable place) so that features are extracted from an observation only once. It might goes beyond CLandmarksMap, e.g. by storing results of any time consuming processes on raw data near themselves. I'm not very familiar with the code to propose a good solution :P
Thanks
