How does Lowe compute the "repeatability" of his SIFT Algorithm?

Computer Vision: How does the SIFT decriptor becomes the drawn vector on the image?

• I run sift algorithm in a image; the output is a variable des, which contains n vectors representing keypoints. I know, from the Lowe paper, that each value of the 128 numbers in the descriptor represents a gradient, but how does it turns into the vector drawn in the image?  How can I manipulate these vectors as to threshold then, or to select the biggest ones? (preferable in python) Thanks!

• Answer:

  The 128-dimensional vector just gives you the feature weights. The row above the feature vector in Lowe's program gives you the x,y coordinates and the x,y orientation of the corresponding feature vector.