some questions about GFLOPs and FPS

[aluds123]

Hi, @PingoLH :
Thanks for all your Projects. I have tested your model on Jetson Nano and it works fast and accurate.
I have some questions:

1. Can I ask how do you calculate GFLOPs?
2. CenterNet-HarDNet85 is the most fast and accurate model in the tabel "Comparison with other state-of-the-art works"
   
   , but why CenterNet-HarDNet85's FPS is similar to YOLOv4 when they run on Jetson Nano?
   https://github.com/jkjung-avt/tensorrt_demos:Demo #5

(Maybe some details I didn’t notice.)
Thank you.

[PingoLH]

Hi,

Thanks for the feedback. GFLOPs was calculated by https://github.com/sovrasov/flops-counter.pytorch.
For 2, I think is because HarDNet still has a lot of concatenate operations, which still couldn't be optimized by TensorRT. Another minor difference is that HarDNet for Jetson was converted through onnx, while YOLOv4 for Jetson was built directly by DarkNet, which might give YOLOv4 a better optimization.

[aluds123]

Hi, @PingoLH :
Thanks for your fast reply. I'll try to understand TensorRT's optimizations.
Thank you very much.