choosing a threshold value to compute the similarity between two images in matlab

Question

In order to compute the distance (D) between the query image feature with each image database feature and display (n-similarity images to the user).

I try to use the following approach:

I choose two threshold values (T1, T2). For the first one I keep all the distances (D2) which are larger than (T1) in a variable say (L1), and keep all the distances (D2) which are smaller than (T2) in another variable, say (L2). Then, I compute the similarity measure by:

S(i) = L2 * average(D3) / (L3^2)

Please, how could I choose these thresholds? Is there is any method to compute the threshold value or should I choose it randomly?

Answer 1

I have trouble understanding your expressions. What exactly is D3? I assume the index i in S(i) refers to the i'th image in the database. Is D or L indexed by i also? In general for problems like these, what is more important is choosing the right similarity measure, and then comparing different approaches with methods like the ROC and Precision-Recall curves. You should leave out worrying about the threshold until then.