Ascend OpenPose

Requirement	Version
OS	Ubuntu 18.04
OpenCV	4.x compiled with GCC 7.5
OpenMP	default with GCC
CANN	3.3.0 or above

supported Ascend Hardware：Atlas300-3000/3010、Atlas800-3000/3010

Performance：

Hardware	e2e latency	model input geometry	FPS
Ascend310	45.4ms	368x448	22

1.Caffe model trans to ascend model

OpenPose body25 model can be download here , passwd:123456

trans yolov5.onnx to om with ATC

source /usr/local/Ascend/ascend-toolkit/set_env.sh
atc --model=body_25.prototxt --weight=body_25.caffemodel --framework=0 --output=body25 --soc_version=Ascend310 --insert_op_conf=aipp_rgb.cfg --input_format=NCHW --input_shape="images:1,3,368,448"

param	statement
--model	prototxt
--weight	caffemodel
--output	output om file
--input_shape	model input shape: (batch, channel, width, height)

2.Compile and Run program

2.1 modify CMakeLists.txt

set(OpenCV_DIR /root/opencv4.4/lib/cmake/opencv4) #your own opencv path
...
set(ACL_PATH /usr/local/Ascend/ascend-toolkit/latest) #your own cann-toolkit path

2.2 Compile it

mkdir build
cd build
cmake ..
make -j8

2.3 Run it

./ACL_OPENPOSE model/body25.om test.mp4

===============================
ModelProcess:Begin to inference.
detected people:5
fps:21.1667
===============================

Reference:

EasyOpenPose: https://github.com/dlunion/EasyOpenPose