search

jwwangchn / NWD

Official code for "A Normalized Gaussian Wasserstein Distance for Tiny Object Detection"
Apache License 2.0
204 stars 23 forks source link

where to find "supplementary" materials？ #3

Open zimenglan-sysu-512 opened 2 years ago

zimenglan-sysu-512 commented 2 years ago

hi @jwwangchn where to find "supplementary" materials? since i want to use nwd to train my own dataset, so that i need to calculate the constant variable C. thanks.

urbaneman commented 2 years ago

I have the same question，in paper Eq.8

". In the following experiments, we empirically set C to the average absolute size of AI-TOD and achieve the best performance. Moreover, we observe that C is robust in a certain range, details will be shown in supplementary materials."

I would like to check the supplementary materials on C to confirm the related experiments of hyperparametric C.

Could you show the relevant experimental results？@jwwangchn

Chasel-Tsui commented 2 years ago

If the tested dataset contains many tiny objects (smaller than 16*16), I think C might be robust from 12 to 16.

ReusJeffery commented 11 months ago

I have the same question，in paper Eq.8

". In the following experiments, we empirically set C to the average absolute size of AI-TOD and achieve the best performance. Moreover, we observe that C is robust in a certain range, details will be shown in supplementary materials."

I would like to check the supplementary materials on C to confirm the related experiments of hyperparametric C.

Could you show the relevant experimental results？@jwwangchn

I found that the C equals 12.8 in this paper is the absolutely size of AITOD dataset,and I just read the paper of AITOD,the paper mentioned that the way to calculate the absolutely size is sqrt(w*h).So you can calculate the C on your own dataset for every instance and add them up and average