While experimenting with TLT retraining of pretrained detectnetv2 resnet10 model for 5 class object detection, I generally find that when the training dataset has large scale objects in it, the final model throws a lot of false positives. Is it a known issue? and how to resolve it ?
There is not a known issue for your case. You can tune the hyperparameters or use larger backbones to check again. Or to check if training dataset is enough, etc. You can also shed light on me how to reproduce your case too.