What’s the fastest model available in TLT?

What’s the fastest model and backbone available in TLT based on FPS performance? Is DetectNet_v2 ResNet10? I want to use it in the DeepStream SDK on the Jetson Nano board.

I’ve trained for DetectNet_v2 ResNet10, but I would like to know if it’s the fastest model without having to train another one (because I don’t have a powerful hardware to do this quickly).


Detectnet_v2 can have a better fps. But please note that fps depends on different training size, backbones, pruning ratio, etc.

DetectNet_v2 + ResNet10 is the best choice? I’m training a small network.

Please note that we need to balance between fps and mAP.

I understand, but I’m getting 90+ mAP with ResNet10 pruned with 0.7 pth. I would like to know if using another model I can get faster inference than ResNet10.

You can trigger experiments to train different pruned model of smaller pruning ratio.