deepstream_yolo

Deploy YOLO Models With DeepStream

This sample shows how to integrate YOLO models with customized output layer parsing for detected objects with DeepStreamSDK.

1. Sample contents:

• deepstream_app_config_yolo.txt: DeepStream reference app configuration file for using YOLO models as the primary detector.
• config_infer_primary_yoloV4.txt: Configuration file for the GStreamer nvinfer plugin for the YoloV4 detector model.
• config_infer_primary_yoloV7.txt: Configuration file for the GStreamer nvinfer plugin for the YoloV7 detector model.
• nvdsinfer_custom_impl_Yolo/nvdsparsebbox_Yolo.cpp: Output layer parsing function for detected objects for the Yolo models.

2. Pre-requisites:

2.1 Please make sure DeepStream 6.1.1+ is properly installed

2.2 Generate Model

YoloV4

• Go to this pytorch repository https://github.com/Tianxiaomo/pytorch-YOLOv4 where you can convert YOLOv4 Pytorch model into ONNX
• Other famous YOLOv4 pytorch repositories as references:
  • https://github.com/WongKinYiu/PyTorch_YOLOv4
  • https://github.com/bubbliiiing/yolov4-pytorch
  • https://github.com/maudzung/Complex-YOLOv4-Pytorch
  • https://github.com/AllanYiin/YoloV4
• Or you can download reference ONNX model directly from here (link).

YOLOv7

following the guide https://github.com/WongKinYiu/yolov7#export, export a dynamic-batch-1-output onnx-model

$ python export.py --weights ./yolov7.pt --grid --simplify --topk-all 100 --iou-thres 0.65 --conf-thres 0.35 --img-size 640 640 --dynamic-batch

or using the qat model exported from yolov7_qat

3. Download and Run

  $ cd ~/
  $ git clone https://github.com/NVIDIA-AI-IOT/yolo_deepstream.git
  $ cd ~/yolo_deepstream/deepstream_yolo/nvdsinfer_custom_impl_Yolo
  $ make
  $ cd ..

Make sure the model exists under ~/yolo_deepstream/deepstream_yolo/. Change the "config-file" parameter in the "deepstream_app_config_yolo.txt" configuration file to the nvinfer configuration file for the model you want to run with.

Model	Nvinfer Configuration File
YoloV4	config_infer_primary_yoloV4.txt
YoloV7	config_infer_primary_yoloV7.txt

  $ deepstream-app -c deepstream_app_config_yolo.txt

4. CUDA Post Processing

this sample provide two ways of yolov7 post-processing(decoce yolo result, not include NMS), CPU version and GPU version

• CPU implement can be found in: nvdsparsebbox_Yolo.cpp
• CUDA implement can be found in: nvdsparsebbox_Yolo_cuda.cu

Default will use CUDA-post processing. To enable CPU post-processing: in config_infer_primary_yoloV7.txt

• parse-bbox-func-name=NvDsInferParseCustomYoloV7_cuda -> parse-bbox-func-name=NvDsInferParseCustomYoloV7
• disable-output-host-copy=1 -> disable-output-host-copy=0

The performance of the CPU-post-processing and CUDA-post-processing result can be found in Performance