I made a multi-scale Harris corner detector (inspired by MOPS) which is simply a Harris corner detector performed on several scales (i.e. subsampled versions of the same image).
This effectively make the detector more robust to noisy/blurred images and able to find more features.
However, there are cases where two or more features are detected at the same location on different scales. See the image - each circle represents an oriented corner, radius of the circle corresponds with its scale:
Because of this, some features have similar descriptors and this would theoretically weaken later correspondence matching.
Detectors like SIFT or Harris-Laplace deal with the problem by either detecting scale-space maxima only or finding characteristic scale of each detected corner.
I think it would be reasonable to group corners that are close to each other spatially and then choose just the strongest one (the one with the strongest response is nearest to its characteristic scale). I am already performing Adaptive Non-Maximal Suppression, but this happens on each scale separately (to keep moderate amount of features on each scale) so the corners with similar locations can still remain when they are on different scales.
Maybe just operating on descriptor vectors would be sufficient, but this problem have much higher dimensionality.