OpenVINO-YOLOV4

Introduction

This is the implementation of YOLOV4,YOLOV4-relu,YOLOV4-tiny in OpenVINO2020R4 .

This update has replaced all v3 interfaces name with v4. In addition to the JSON file: "id": "TFYOLOV3"(demo code needs to be modified if "id": "TFYOLOV4 is used) .we want to make sure that you can run the demo on your PC directly

Environment

OpenVINO2020R4 :https://docs.openvinotoolkit.org/latest/index.html

Win or Ubuntu

YOLOV4:https://github.com/AlexeyAB/darknet and download weights file

*Convert YOLOV3/2/1 model :https://docs.openvinotoolkit.org/latest/openvino_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html

How to use

YOLOV4

download yolov4.weights .

#windows  default OpenVINO path

python convert_weights_pb.py --class_names cfg/coco.names --weights_file yolov4.weights --data_format NHWC

"C:\Program Files (x86)\IntelSWTools\openvino\bin\setupvars.bat"

python "C:\Program Files (x86)\IntelSWTools\openvino_2020.4.287\deployment_tools\model_optimizer\mo.py" --input_model frozen_darknet_yolov4_model.pb --transformations_config yolov4.json --batch 1

python object_detection_demo_yolov3_async.py -i cam -m frozen_darknet_yolov4_model.xml  -d CPU

This is OpenVINO2020R4 object_detection_demo_yolov3_async.py without any change,but objects can still be detected normally

CPU(intel i5-8250U)

YOLOV4-relu

download yolov4.weights .

#windows  default OpenVINO path
cd yolov4-relu

python convert_weights_pb.py --class_names cfg/coco.names --weights_file yolov4.weights --data_format NHWC

"C:\Program Files (x86)\IntelSWTools\openvino\bin\setupvars.bat"

python "C:\Program Files (x86)\IntelSWTools\openvino_2020.4.287\deployment_tools\model_optimizer\mo.py" --input_model frozen_darknet_yolov4_model.pb --transformations_config yolov4.json --batch 1

python object_detection_demo_yolov3_async.py -i cam -m frozen_darknet_yolov4_model.xml  -d CPU

YOLOV4-tiny

download yolov4-tiny.weights .

#windows  default OpenVINO path

python convert_weights_pb.py --class_names cfg/coco.names --weights_file yolov4-tiny.weights --data_format NHWC --tiny

"C:\Program Files (x86)\IntelSWTools\openvino\bin\setupvars.bat"

python "C:\Program Files (x86)\IntelSWTools\openvino_2020.4.287\deployment_tools\model_optimizer\mo.py" --input_model frozen_darknet_yolov4_model.pb --transformations_config yolo_v4_tiny.json --batch 1

python object_detection_demo_yolov3_async.py -i cam -m frozen_darknet_yolov4_model.xml  -d CPU

when you convert .pb to IR,it is difficult to avoid the loss of accuracy(just like yolov3-tiny).if you couldn't detect objects correctly, you could further train yolov4-tiny.weights (https://github.com/AlexeyAB/darknet)

object_detection_demo_yolov4_async.py

(1)Add DIOU-NMS Support

buptdbj/OpenVINO-YOLOV4