juanarocha's profile - activity

Yes that's what I normally do. With 1500 (as is done in OpenCV examples) and then thresholding, but I've seen some posts recommending 300-400 points. So, I was wondering if there's some kind of standard or recommended values. ORB's paper doesn't say anything about it

OpenCV OrbFeatureDetector's detect function has a parameter nFeatures which tells the function what is the maximum number of keypoints to be detected.

I know more points means more comparisons when matching so less performance, while less means less accuracy.

Is there a recommended number of keypoints that gives me a good accuracy/performance relation? And if there is, why?

An ORB descriptor Mat object contains X rows (X depending on the max number of features) which represent the features descriptors and 32 columns in which are stored integers between 0 and 255 which represent the descriptor elements. My assumption is that the binary representation of the descriptor is to change these integers to their binary representation (8 bits each, 8*32=256 which is ORB descriptors size) and concatenate them (please correct me if I'm wrong).

Has OpenCV an automatic way to deal with this? converting the int representation to actually binary?