OpenCV SIFT vs. David Lowe's SIFT implementation, OpenCV is much worse

Question

I tried object detection for several different objects using the implementation of David Lowe and OpenCV (just changed 'SurfFeatureDetector' → 'SiftFeatureDetector' and 'SurfDescriptorExtractor' → 'SiftDescriptorExtractor' in this code).

The opencv implementation seems to be much worse! Found many false positives and fewer key points! I tried to play with SiftFeatureDetector parameters like sigma / contrastThreshold, etc., but the results were always much worse than the original version of David.

• Is the opencv implementation of SIFT completely “different” from the original implementation of David Lowe?

• Does anyone know of a good set of parameters for the 'SiftFeatureDetector', so the results for images of objects equipped with the David implementation (for example, box.pgm / basmati.pgm ..) may seem similar to the version of David.

Answer 1

VLFeat SIFT implementation is pretty close (in terms of function output) to the original implementation.

Answer 2

I use the OpenCV implementation of SIFT and compare it with the David Lowe implementation available on his website. When setting up some parameters, I had good results.