search

NVIDIA / DIGITS

Deep Learning GPU Training System
https://developer.nvidia.com/digits
BSD 3-Clause "New" or "Revised" License
4.12k stars 1.38k forks source link

Detection of objects smaller than 50x50 px #1503

Open hadign20 opened 7 years ago

hadign20 commented 7 years ago

I'm trying to detect vehicles in aerial images using digits. So far I have customized my dataset to be as KITTI format and I changed some size values in the detectNet_network.prototxt to fit the input images. The images are of size 1024x1024 and the objects are mostly smaller than 50x50 pixels. In the training phase, my output looks this way: screenshot from 2017-03-09 21-02-58

As you see, the mAP is on 0 and it's not working. I assume it's because of the objects being small or setting the dimensions in "detectNet_network.prototxt" to 1024x1024 isn't right. My question is, can I use any other architectures other than DetectNet maybe with a different stride (like 8) to solve this problem, or that's not suggested?

brash6 commented 5 years ago

Hi, I have the same problem ! Have you been able to solve this issue ? Thanks