The confidence of each match can be deduced by the descriptors distance. More elaborate mechanisms are the ratio test, fitting a transformation between images and removing outliers via ransac or best buddies similarity.

]]>in the official it says x and y run from [-r, r] I did not get it? how to calculate the r? ]]>

that’s a very nice explanation.

one thing that is missing: “the accuracy”. matching descriptors is great, but how can I tell the confidence of each match? ]]>

Can you explain how the variance between two pairs of binary vectors is calculated ?

]]>I was looking into the opencv implementation ORB.

I have two questions for which I could not find any answers till now.Hope you could help :).

in the orb.cpp file , we could see after calculating the orientation of the patch,new rotated coordinates was computed using sampling pair and angle calculated for the patch.

I have a doubt, in boundary region there might be a patch for which after calculating the angle of the patch and subsequently calculating the coordinates ,the pixels to be selected might lie outside the frame itself. How we need to handle it in ORB?

My second doubt is ,while calculating the centre pixel address why layer.x and layer.y has been added what are the importance of that.

I need to implement ORB completely on embedded hardware so i need to understand the code thoroughly before I could implement the code on embedded board.

Thanks in advance 🙂

Arun