Running YOloV4 on jetson Nano at Higher FPS?

Hi

We are trying to Run Yolov4 on jetsonNano developer kit 4gb Ram, but So far we have only managed to get 1Fps we need at least 4fps.

is this because we are building it wrong for jetson?

THe yolov4 model was build based on this repo https://github.com/AlexeyAB/darknet

We are new to jetson Nano, Anyone who knows a way or little guidance it would help us a lot

We can use yolov3 or tiny yolo for higher fps but we can not compromise on accuracy, So Please help us

Thank you

Hi,

Usually, we can get a better performance with Deepstream since it applied TensorRT acceleration.

Not sure if DeepstreamSDK sample can fully support YOLOv4.
Would you mind to give it a try first?

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

Thanks.

1 Like

I have implemented TensorRT YOLOv4. These are the performance numbers I got when I tested it on Jetson Nano with JetPack-4.4.

TensorRT engine mAP @
IoU=0.5:0.95
mAP @
IoU=0.5
FPS on Nano
yolov4-288 (FP16) 0.372 0.590 6.18
yolov4-416 (FP16) 0.454 0.698 3.50
yolov4-608 (FP16) 0.484 0.735 1.77
1 Like

Recently AlexeyAB released YOLOv4-Tiny. It has been indicated to yield 39 FPS on the nano and 290 FPS on the AGX for 416x416. Have a look at the official repo at https://github.com/AlexeyAB/darknet

I’ve had some interesting discussion with AlexeyAB about TensorRT yolov4 and yolov4-tiny FPS numbers on Jetson Nano. Check it out here: https://github.com/AlexeyAB/darknet/pull/5453#issuecomment-663593512

Here’s a quick update of FPS numbers (on Jetson Nano) after I updated my tensorrt yolov4 implementation with a “yolo_layer” plugin. For example, “yolov4-416” (FP16) has been improved to 4.62 FPS. Refer to my jkjung-avt/tensorrt_demos repository for details.

1 Like

Does deepstream 5.0 support yolov4?

You can use my repo for yolov4 or yolov5 on deepstream

1 Like