Hi Eduardo,
Using a filter can certainly make things worse. We found when using a moving average filter (not the same as a median filter, but similar), we had to "tune" the window size. Too small resulted in little impact, too large distorted measurements too much. To give you an idea, running an Aspirin 1.5 on a TWOG over i2c with a periodic frequency of 120Hz with a window of 4 gave barely noticeable improvements, a window of 8 gave fairly nice performance, and a window of 12 started to give wonky results and behaviour. In addition, we found that filtering gyro data made performance poorer, while just filtering accelerometer data helped. We also noticed differences in the frequencies that affected the gyro and the accelerometer, and each axis was also a little bit different. Definitely take a look at each sensor axis across the whole throttle range, it might provide a bit of insight.
You may have to experiment a bit...
(The moving average filter code is not the nicest, but it is here if you want to look, in the imu_ppzuav.[h|c] files:
Thanks,
-Stephen Dwyer