1 // Tencent is pleased to support the open source community by making ncnn available.
2 //
3 // Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
4 //
5 // Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
6 // in compliance with the License. You may obtain a copy of the License at
7 //
8 // https://opensource.org/licenses/BSD-3-Clause
9 //
10 // Unless required by applicable law or agreed to in writing, software distributed
11 // under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
12 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
13 // specific language governing permissions and limitations under the License.
14
15 #include "net.h"
16
17 #include <algorithm>
18 #include <opencv2/core/core.hpp>
19 #include <opencv2/highgui/highgui.hpp>
20 #include <opencv2/imgproc/imgproc.hpp>
21 #include <stdio.h>
22 #include <vector>
23
24 struct KeyPoint
25 {
26 cv::Point2f p;
27 float prob;
28 };
29
detect_posenet(const cv::Mat & bgr,std::vector<KeyPoint> & keypoints)30 static int detect_posenet(const cv::Mat& bgr, std::vector<KeyPoint>& keypoints)
31 {
32 ncnn::Net posenet;
33
34 posenet.opt.use_vulkan_compute = true;
35
36 // the simple baseline human pose estimation from gluon-cv
37 // https://gluon-cv.mxnet.io/build/examples_pose/demo_simple_pose.html
38 // mxnet model exported via
39 // pose_net.hybridize()
40 // pose_net.export('pose')
41 // then mxnet2ncnn
42 // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models
43 posenet.load_param("pose.param");
44 posenet.load_model("pose.bin");
45
46 int w = bgr.cols;
47 int h = bgr.rows;
48
49 ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, w, h, 192, 256);
50
51 // transforms.ToTensor(),
52 // transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
53 // R' = (R / 255 - 0.485) / 0.229 = (R - 0.485 * 255) / 0.229 / 255
54 // G' = (G / 255 - 0.456) / 0.224 = (G - 0.456 * 255) / 0.224 / 255
55 // B' = (B / 255 - 0.406) / 0.225 = (B - 0.406 * 255) / 0.225 / 255
56 const float mean_vals[3] = {0.485f * 255.f, 0.456f * 255.f, 0.406f * 255.f};
57 const float norm_vals[3] = {1 / 0.229f / 255.f, 1 / 0.224f / 255.f, 1 / 0.225f / 255.f};
58 in.substract_mean_normalize(mean_vals, norm_vals);
59
60 ncnn::Extractor ex = posenet.create_extractor();
61
62 ex.input("data", in);
63
64 ncnn::Mat out;
65 ex.extract("conv3_fwd", out);
66
67 // resolve point from heatmap
68 keypoints.clear();
69 for (int p = 0; p < out.c; p++)
70 {
71 const ncnn::Mat m = out.channel(p);
72
73 float max_prob = 0.f;
74 int max_x = 0;
75 int max_y = 0;
76 for (int y = 0; y < out.h; y++)
77 {
78 const float* ptr = m.row(y);
79 for (int x = 0; x < out.w; x++)
80 {
81 float prob = ptr[x];
82 if (prob > max_prob)
83 {
84 max_prob = prob;
85 max_x = x;
86 max_y = y;
87 }
88 }
89 }
90
91 KeyPoint keypoint;
92 keypoint.p = cv::Point2f(max_x * w / (float)out.w, max_y * h / (float)out.h);
93 keypoint.prob = max_prob;
94
95 keypoints.push_back(keypoint);
96 }
97
98 return 0;
99 }
100
draw_pose(const cv::Mat & bgr,const std::vector<KeyPoint> & keypoints)101 static void draw_pose(const cv::Mat& bgr, const std::vector<KeyPoint>& keypoints)
102 {
103 cv::Mat image = bgr.clone();
104
105 // draw bone
106 static const int joint_pairs[16][2] = {
107 {0, 1}, {1, 3}, {0, 2}, {2, 4}, {5, 6}, {5, 7}, {7, 9}, {6, 8}, {8, 10}, {5, 11}, {6, 12}, {11, 12}, {11, 13}, {12, 14}, {13, 15}, {14, 16}
108 };
109
110 for (int i = 0; i < 16; i++)
111 {
112 const KeyPoint& p1 = keypoints[joint_pairs[i][0]];
113 const KeyPoint& p2 = keypoints[joint_pairs[i][1]];
114
115 if (p1.prob < 0.2f || p2.prob < 0.2f)
116 continue;
117
118 cv::line(image, p1.p, p2.p, cv::Scalar(255, 0, 0), 2);
119 }
120
121 // draw joint
122 for (size_t i = 0; i < keypoints.size(); i++)
123 {
124 const KeyPoint& keypoint = keypoints[i];
125
126 fprintf(stderr, "%.2f %.2f = %.5f\n", keypoint.p.x, keypoint.p.y, keypoint.prob);
127
128 if (keypoint.prob < 0.2f)
129 continue;
130
131 cv::circle(image, keypoint.p, 3, cv::Scalar(0, 255, 0), -1);
132 }
133
134 cv::imshow("image", image);
135 cv::waitKey(0);
136 }
137
main(int argc,char ** argv)138 int main(int argc, char** argv)
139 {
140 if (argc != 2)
141 {
142 fprintf(stderr, "Usage: %s [imagepath]\n", argv[0]);
143 return -1;
144 }
145
146 const char* imagepath = argv[1];
147
148 cv::Mat m = cv::imread(imagepath, 1);
149 if (m.empty())
150 {
151 fprintf(stderr, "cv::imread %s failed\n", imagepath);
152 return -1;
153 }
154
155 std::vector<KeyPoint> keypoints;
156 detect_posenet(m, keypoints);
157
158 draw_pose(m, keypoints);
159
160 return 0;
161 }
162