The result must be tested using the "Bull's Eye Test": for every shape, compute the top 40 similar shapes out of the entire (1400) database. Then count how many hits (=same class) you get, max. is 20 (since there are only 20 shapes in each class). Average the number over all shapes and compute the percentage (20 = 100%). Remark: since it should be easy to detect the exact same shape (and the query is amongst the 1400 target shapes), a hit percentage of 5% (1 out of 20) is a fairly minimal success...
The approach is old (1996), but it is close to as good as it gets, if you constrain yourself to 2D boundaries and purely geometric approaches (no machine learning in there).
The relevant papers are these two:
Database (plus utility MATLAB functions) is HERE.Good luck and enjoy!