DIGITS icon indicating copy to clipboard operation
DIGITS copied to clipboard

Published 20 hours ago • verified publisher icon NVIDIA

Is Stride Adjustable?

Open jtrovato opened this issue 7 years ago • 2 comments

Hello,

I am attempting to DetectNet to detect smaller objects (15-100px). From what I have read the default stride of 16 is better for larger onbjects (50-400px). I am trying to adjust the stride but adjusting the stride seem to change the dimensionality of a few different layers, causing the training to error. I have traced the net back and it seem that the net is specifically tailored to a stride of 16. Is there a procedure for lowering the stride in the DetectNet implementation?

jtrovato avatar Oct 10 '17 17:10 jtrovato

Hey,

I have been successfully able to train for smaller objects. Here are the links that helped me:

https://groups.google.com/d/topic/digits-users/zx_UYu3jlt8/discussion https://github.com/NVIDIA/DIGITS/issues/1521 https://github.com/NVIDIA/DIGITS/issues/980 https://github.com/NVIDIA/DIGITS/issues/1011

There are a few more issues if you can dig a little bit...

Hope it helps.

kanakiyab avatar Oct 10 '17 20:10 kanakiyab

Hey kanakiyab Could you share a little more details about how you done it? And can you achieve high accuracy with small objects (like 15-20 px) ?

Thanks in advance

kadirbeytorun avatar Jun 11 '19 11:06 kadirbeytorun