Tuning SIFT algorithm parameters?

asked 2016-12-23 15:16:05 -0600

lovaj gravatar image

I'm trying to produce VLAD codes starting from SIFT descriptors using OpenCV implementation with the Oxford Building Dataset.

SIFT has many parameters such as #octaves, #scales, contrastThreshold and edgeThreshold. The right values for these parameters can significantly improve descriptors VLAD precision, so tuning is needed.

What is a good way for this task? Any particular procedure is present? Any suggestion is more than welcome!

edit retag flag offensive close merge delete

Comments

Parameter Optimization. For this you'll need test and validation data and some way of quantifying "improvement" or "goodness", and once you identify that, you then need to write a program to search for the best parameters. This can be random or exhaustive or based on gradient. I think it's probably a fun exercise, best of luck!

Der Luftmensch gravatar imageDer Luftmensch ( 2016-12-23 15:27:20 -0600 )edit

"improvement" is given by a better mean average precision on the final generated VLAD codes using the considered dataset. I've never tuned an algorithm like that, someone suggested me cross validation.

lovaj gravatar imagelovaj ( 2016-12-24 00:33:32 -0600 )edit