What I should do about BSDS300
Post date: Sep 21, 2011 7:58:34 PM
What I should do about BSDS300
- Pick the level from UCM that I want to use. Pick one with F = 0.68. Let's call it UCM_a
- Try to beat UCM_a with DDT.
Evaluate using BSDS
UCM experiment
Mori_A takes
0.21, 0.29, 0.67, 1.19, 2.83, 3.55 minutes for 5, 20, 60, 120, 240, 300 superpixels respectively.
*** Observe that the P (0.67-->0.71) is higher and R (0.74-->0.68) is lower after DDT is applied to the segmentation. Therefore, I have an idea to preserve score by adding the UCM_D back to the result, that is, the final contour C = DDT(UCM_D) + UCM_D. I think this approach would increase F-measure.
Waiting list for BSDS evaluation