YOLOV3 frame rate on Xavier

Hi, I have been using the c code implementation of YOLOV3 from https://github.com/pjreddie/darknet. It is a complete c/cuda implementation without using any TensorFlow, Caffee stuff. I trained my own tiny" model and could achieve about 40 FPS on 640x360 image on Xavier. However, today, I watched a talk given by Dustin Franklin and he showed YOLOV3 could achieve 1000FPS on Xavier! He mentioned something like TLT or maybe ontop of Tensor RT. I would like to know how can I benefit from TLT and TensorRT to achieve the same throughtput. Can I still use the github c implemenation or I need to use TensofFlow/Caffee etc? Please note I need to train my own model with my own data. I can not use Nvidia pre-trained model.

Hi,

Would you mind to share the video you mentioned so we can redirect the corresponding link to you?
In general, we recommend to use Deepstream for the YOLO optimization.

Our Deepstream sample optimized YOLO directly from darknet model format which should meet your requirement.

/opt/nvidia/deepstream/deepstream-5.0/sources/objectDetector_Yolo/

Thanks.

Hi,

You can check this sample for the entire pipeline for YOLO model:

/opt/nvidia/deepstream/deepstream-4.0/sources/objectDetector_Yolo/

OR this sample for the inference part benchmarking:

Thanks.

Where can I get the TensorRT YOLOV3 code?

@AutoCar,

https://docs.nvidia.com/deeplearning/tensorrt/sample-support-guide/index.html

Is it what you need?
you can find OD with TensorRT.

This is the onnx implementation. I am using the author’s original c implementation. I am not sure my trained mode can use used here.

Hi,

You can check this sample for the entire pipeline for YOLO model:

/opt/nvidia/deepstream/deepstream-4.0/sources/objectDetector_Yolo/

OR this sample for the inference part benchmarking:

Thanks.