Epipolar geometry check for feature tracking?

[steevo87]

Previously, for filterVectors() you would set the epipolarCheck variable as true. However, this was recognized to remove good features when the epipolar test was invalid, e.g. when the features were all clustered, or there were very few of them, and the RANSAC-based method was ineffective in determining a good epipolar estimate. Do you think it is worthwhile making setting a minimum number of features and/or minimum feature spread for the epipolar check? And also, is it necessary to add it as an optional configurable variable?

There is still a mystery though.. regarding the test in the Downloads\minh_bug directory. With the epipolar check on, you get a different preserved set of points using Release than Debug. In either build, it's the same set each time! It seems to match Minh's, as well, so not computer dependent. Maybe it is something to do with the random selection of points for RANSAC differing between Release and Debug.