0 Loss and AP

Question

0 Loss and AP

alfinnurhalim opened this issue 2 years ago · comments

Hi. Thank you for your great work! I have successfully trained and tested your model using KITTI dataset and it works.

I currently trying to train a custom dataset (which is not car), but somehow the loss and AP is 0

I have correctly set the image size and the dataset was already validated and it is good. do you have any suggestion? I might miss some config.

Alfin Nurhalim · Answer 1 · Wed Dec 22 2021 22:23:00 GMT+0800 (China Standard Time)

One more thing, how did you calculate the voxel size? thank you in advance

Danila Rukhovich · Answer 2 · Wed Dec 22 2021 22:52:01 GMT+0800 (China Standard Time)

It's hard to help you, as I don't know your dataset. Looks like the loss is zero. So I think none of your objects are assigned to anchors. You need to check anchor ranges, anchor sizes, origin.

We use voxel size of 0.32m to be the same as a default distance between anchors for point cloud methods on KITTI.

Alfin Nurhalim · Answer 3 · Wed Dec 22 2021 23:18:00 GMT+0800 (China Standard Time)

I see. thank you for the info. let me check it then

Alfin Nurhalim · Answer 4 · Thu Dec 23 2021 19:24:18 GMT+0800 (China Standard Time)

What does n_voxels denotes to? how did you calculate it? if you got this from other paper, would you please refer which one?

I'm currently studying your paper, thank you for all your help

Danila Rukhovich · Answer 5 · Fri Dec 24 2021 02:52:52 GMT+0800 (China Standard Time)

n_voxels is about N_x, N_y, N_z from the paper.

Alfin Nurhalim · Answer 6 · Mon Dec 27 2021 21:18:19 GMT+0800 (China Standard Time)

Hi, thank you for your answer. I still have several question regarding your paper. I will close this issue and write you an email instead for more detail. thank you

Alfin Nurhalim · Answer 7 · Mon Dec 27 2021 21:19:59 GMT+0800 (China Standard Time)

I'll send to your email written in the paper, is it good?, let me know if you change your email. thank you