search

princeton-vl / relative_depth

Code for the NIPS 2016 paper
BSD 3-Clause "New" or "Revised" License
124 stars 31 forks source link

Problem with WKDR on NYU dataset #4

Closed ChenyangLiu closed 6 years ago

ChenyangLiu commented 6 years ago

Hi,

I noticed that in your paper, Table 2, your WKDR is less than both WKDR_eq and WKDR_neq. But I think that, WKDR should be a weighted ave of them. How do you get the WKDR?

wfchen-umich commented 6 years ago

Hi, thank you for your interest in our paper. Please take a look at: https://github.com/umich-vl/relative_depth/blob/master/src/experiment/test_model_on_NYU.lua

ChenyangLiu commented 6 years ago

I've read this code, and found that your WKDR is 1-(n_eq+n_neq)/(gt_eq+gt_neq), it should be between 1-n_eq/gt_eq and 1-n_neq/gt_neq, which are WKDR_eq and WKDR_neq. But in the paper, WKDR in less than both WKDR_eq and WKDR_neq. That makes me confused.

wfchen-umich commented 6 years ago

I think that is because we take the mean of WKDR from each image, but not calculating the WKDR of all the pairs as a whole.