Multimodal landmarks for EKF

edit

I am trying to do a robot localization using visual landmarks that are not so robust, which means that at each observed landmark, there is a not so small probability that it could be any of two or three similar landmarks that were observed at different places in my map. This could result in eventual jumps in the robot's estimated pose (similar to AMCL). The theorz says that Kalman filter provide an optimal solution when the observations are unimodal, but in practice can a localization based on this kind of data combined with odometry and using an extended Kalman filter still work?