Issue with detecting small objects using Detectnet

[Coldcammey]

@dusty-nv
Hello there. I am having a recurrent issue regarding the detectnet.
I am training using a custom dataset with ssd-mobilenet, trying to train a traffic speed limit sign dataset. I train the network on my computer, then use it on my 4gb Jetson-Nano.

My issue is that no matter what I do, I cannot make it detect smaller objects. No matter if it's 10 or 80 epochs, 16 or 64 batch_sizes, 1 or more workers, I cannot make it detect smaller objects. Even if the object is close, the performance of the detect is really poor. I've used the 512x512 version as well. But saw very little to no improvements.

Some examples:

I don't know what is the issue. I have followed the github and youtube tutorials I've changed so many datasets so I don't think the datasets are the issue here. I try to keep as least 100-150 images per class if not more.

I would be glad if anyone can help me.