How do we determine what is a good correspondence and what is a bad correspondence between the matched feature points? (other than us visually determining that)
In short, the goal of RANSAC is to find model parameters that best represent how the majority of feature locations are transformed from one image to the other. If many correspondences agree with a given model, then it provides confidence that we chose the right model parameters. Then, inliers would be considered "good correspondences", and outliers would be "bad correspondences". If only a few correspondences agree with a given model, it might suggest that the model parameters are off.
tapioca
ohh okay, I got the process backwards in my head and for some reason thought we knew what points were inliers and what points were outliers beforehand, but this makes so much more sense now. Thank you!
How do we determine what is a good correspondence and what is a bad correspondence between the matched feature points? (other than us visually determining that)
With RANSAC! (See next slide.)
In short, the goal of RANSAC is to find model parameters that best represent how the majority of feature locations are transformed from one image to the other. If many correspondences agree with a given model, then it provides confidence that we chose the right model parameters. Then, inliers would be considered "good correspondences", and outliers would be "bad correspondences". If only a few correspondences agree with a given model, it might suggest that the model parameters are off.
ohh okay, I got the process backwards in my head and for some reason thought we knew what points were inliers and what points were outliers beforehand, but this makes so much more sense now. Thank you!