HRnetv2 pytorch—>onnx—>opencv推理
一、下载模型、克隆项目
在mmpose官方下载https://mmpose.readthedocs.io/zh_CN/latest/topics/hand%282d%29.html
项目克隆：
git cloen https://github.com/open-mmlab/mmpose
需要的库支持：
onnx
onnxruntime
mmpose
mmcv
二、模型转换pytorch—>onnx
在mmpose根目录下：
python tools/deployment/pytorch2onnx.py configs/hand/2d_kpt_sview_rgb_img/topdown_heatmap/coco_wholebody_hand/hrnetv2_w18_coco_wholebody_hand_256x256.py /workspace/downloads/hrnetv2_w18_coco_wholebody_hand_256x256-1c028db7_20210908.pth --output-file hrnetv2_w18_coco_wholebody_hand_256x256.onnx
```
](https://upload-images.jianshu.io/upload_images/15646173-eab84a4374502725.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
](https://upload-images.jianshu.io/upload_images/15646173-7461750f4f57fdf2.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
](https://upload-images.jianshu.io/upload_images/15646173-df22f88d988833fa.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)
会有警告，为了避免出现其他问题，这里参照网上：
```bash
python remove_initializer_from_input.py --input your_old_model.onnx --output your_new_model.onnx
remove_initializer_from_input.py代码如下：
import onnx
import argparse
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--input", required=True, help="input model")
parser.add_argument("--output", required=True, help="output model")
args = parser.parse_args()
return args
def remove_initializer_from_input():
args = get_args()
model = onnx.load(args.input)
if model.ir_version < 4:
print(
'Model with ir_version below 4 requires to include initilizer in graph input'
)
return
inputs = model.graph.input
name_to_input = {}
for input in inputs:
name_to_input[input.name] = input
for initializer in model.graph.initializer:
if initializer.name in name_to_input:
inputs.remove(name_to_input[initializer.name])
onnx.save(model, args.output)
if __name__ == '__main__':
remove_initializer_from_input()
三、opencv推理代码
#include <opencv2/dnn.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
using namespace cv;
using namespace cv::dnn;
#include <iostream>
using namespace std;
// connection table, in the format [model_id][pair_id][from/to]
// please look at the nice explanation at the bottom of:
// https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/output.md
//
const int POSE_PAIRS[3][20][2] = {
{ // COCO body
{1,2}, {1,5}, {2,3},
{3,4}, {5,6}, {6,7},
{1,8}, {8,9}, {9,10},
{1,11}, {11,12}, {12,13},
{1,0}, {0,14},
{14,16}, {0,15}, {15,17}
},
{ // MPI body
{0,1}, {1,2}, {2,3},
{3,4}, {1,5}, {5,6},
{6,7}, {1,14}, {14,8}, {8,9},
{9,10}, {14,11}, {11,12}, {12,13}
},
{ // hand
{0,1}, {1,2}, {2,3}, {3,4}, // thumb
{0,5}, {5,6}, {6,7}, {7,8}, // pinkie
{0,9}, {9,10}, {10,11}, {11,12}, // middle
{0,13}, {13,14}, {14,15}, {15,16}, // ring
{0,17}, {17,18}, {18,19}, {19,20} // small
} };
int main(int argc, char **argv)
{
CommandLineParser parser(argc, argv,
"{ h help | false | print this help message }"
"{ p proto | | (required) model configuration, e.g. hand/pose.prototxt }"
"{ m model | | (required) model weights, e.g. hand/pose_iter_102000.caffemodel }"
"{ i image | | (required) path to image file (containing a single person, or hand) }"
"{ d dataset | | specify what kind of model was trained. It could be (COCO, MPI, HAND) depends on dataset. }"
"{ width | 368 | Preprocess input image by resizing to a specific width. }"
"{ height | 368 | Preprocess input image by resizing to a specific height. }"
"{ t threshold | 0.1 | threshold or confidence value for the heatmap }"
"{ s scale | 0.003922 | scale for blob }"
);
// cv::String modelTxt = samples::findFile(parser.get<string>("proto"));
// cv::String modelBin = samples::findFile(parser.get<string>("model"));
// cv::String imageFile = samples::findFile(parser.get<String>("image"));
// cv::String dataset = parser.get<cv::String>("dataset");
// int W_in = parser.get<int>("width");
// int H_in = parser.get<int>("height");
// float thresh = parser.get<float>("threshold");
// float scale = parser.get<float>("scale");
cv::String modelOnnx = "C:\\Users\\haihan\\Desktop\\new_hrnetv2_w18_coco_wholebody_hand_256x256.onnx";
cv::String modelTxt = "E:\\openpose_pose_coco.prototxt";
cv::String modelBin = "E:\\code\\openpose-1.7.0\\models\\pose\\coco\\pose_iter_440000.caffemodel";
cv::String imageFile = "E:\\code\\openpose-1.7.0\\examples\\media\\1.png";
cv::String dataset = "COCO";
int W_in = 256;
int H_in = 256;
float thresh = 0.1f;
float scale = 0.003922f;
if (parser.get<bool>("help") || modelTxt.empty() || modelBin.empty() || imageFile.empty())
{
cout << "A sample app to demonstrate human or hand pose detection with a pretrained OpenPose dnn." << endl;
parser.printMessage();
return 0;
}
int midx, npairs, nparts;
if (!dataset.compare("COCO")) { midx = 0; npairs = 17; nparts = 18; }
else if (!dataset.compare("MPI")) { midx = 1; npairs = 14; nparts = 16; }
else if (!dataset.compare("HAND")) { midx = 2; npairs = 20; nparts = 22; }
else
{
std::cerr << "Can't interpret dataset parameter: " << dataset << std::endl;
exit(-1);
}
// read the network model
Net net = readNetFromONNX(modelOnnx);
// and the image
Mat img = imread(imageFile);
if (img.empty())
{
std::cerr << "Can't read image from the file: " << imageFile << std::endl;
exit(-1);
}
Mat img_rgb;
cvtColor(img, img_rgb, COLOR_BGR2RGB);
// send it through the network
Mat inputBlob = blobFromImage(img_rgb, scale, Size(W_in, H_in), Scalar(0, 0, 0), false, false);
net.setInput(inputBlob);
Mat result = net.forward();
// the result is an array of "heatmaps", the probability of a body part being in location x,y
int H = result.size[2];
int W = result.size[3];
// find the position of the body parts
vector<Point> points(22);
for (int n = 0; n < nparts; n++)
{
// Slice heatmap of corresponding body's part.
Mat heatMap(H, W, CV_32F, result.ptr(0, n));
// 1 maximum per heatmap
Point p(-1, -1), pm;
double conf;
minMaxLoc(heatMap, 0, &conf, 0, &pm);
if (conf > thresh)
p = pm;
points[n] = p;
}
// connect body parts and draw it !
float SX = float(img.cols) / W;
float SY = float(img.rows) / H;
for (int n = 0; n < npairs; n++)
{
// lookup 2 connected body/hand parts
Point2f a = points[POSE_PAIRS[midx][n][0]];
Point2f b = points[POSE_PAIRS[midx][n][1]];
// we did not find enough confidence before
if (a.x <= 0 || a.y <= 0 || b.x <= 0 || b.y <= 0)
continue;
// scale to image size
a.x *= SX; a.y *= SY;
b.x *= SX; b.y *= SY;
line(img, a, b, Scalar(0, 200, 0), 2);
circle(img, a, 3, Scalar(0, 0, 200), -1);
circle(img, b, 3, Scalar(0, 0, 200), -1);
}
imshow("OpenPose", img);
waitKey();
return 0;
}
代码为网上的代码，连线部分有问题，需要改进
TO-DO
- HRnetv2 pytorch转onnx
- opencv onnx推理
- opencv dnn cuda加速
