Scaled-YOLOv4 in Deepstream

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU) GPU
• DeepStream Version 5.0.1
• JetPack Version (valid for Jetson only)
• TensorRT Version 7.0
• NVIDIA GPU Driver Version (valid for GPU only) 440
• Issue Type( questions, new requirements, bugs) questions

Hi I have previously run yolov4 on deepstream using https://github.com/NVIDIA-AI-IOT/yolov4_deepstream/blob/master/deepstream_yolov4/README.md and it works well.

I see that recently there is a vastly improved Yolov4 https://github.com/AlexeyAB/darknet/issues/7087 Scaled-Yolov4. I tried to use the new cfg file and weights for this version with the above nvidia deepstream github but while it builds the results are worse than normal yolov4.

Is there something I need to change to be able to use this new model? Thanks

1 Like

Does the Scaled-YoloV4 require different pre-processing, e.g. maintain-aspect-ratio ?

Not sure, reading the paper now https://arxiv.org/pdf/2011.08036.pdf

If it does can I try and set maintain-aspect-ratio=0? I note in yolov4 config its =1

yes, you could take a quick try.

And, you can use https://github.com/NVIDIA-AI-IOT/yolov4_deepstream/tree/master/tensorrt_yolov4 to evaluate the accuracy with Coco dataset.

Hi,

I am also very interested in this topic. Did you finally get the chanve to try it in Deepstream and if so, what was the performance?

I was very surprised as they stated that with Jetson Nano it was possible to get up to 39 FPS, which is a very promising performance.

https://alexeyab84.medium.com/scaled-yolo-v4-is-the-best-neural-network-for-object-detection-on-ms-coco-dataset-39dfa22fa982

Best regards,
Alberto