1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
|
/*
// Copyright (C) 2021-2022 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
*/
#include "models/associative_embedding_decoder.h"
#include <algorithm>
#include <iterator>
#include <limits>
#include <numeric>
#include <vector>
#include <utils/kuhn_munkres.hpp>
void findPeaks(const std::vector<cv::Mat>& nmsHeatMaps,
const std::vector<cv::Mat>& aembdsMaps,
std::vector<std::vector<Peak>>& allPeaks,
size_t jointId,
size_t maxNumPeople,
float detectionThreshold) {
const cv::Mat& nmsHeatMap = nmsHeatMaps[jointId];
const float* heatMapData = nmsHeatMap.ptr<float>();
cv::Size outputSize = nmsHeatMap.size();
std::vector<int> indices(outputSize.area());
std::iota(std::begin(indices), std::end(indices), 0);
std::partial_sort(std::begin(indices),
std::begin(indices) + maxNumPeople,
std::end(indices),
[heatMapData](int l, int r) {
return heatMapData[l] > heatMapData[r];
});
for (size_t personId = 0; personId < maxNumPeople; personId++) {
int index = indices[personId];
int x = index / outputSize.width;
int y = index % outputSize.width;
float tag = aembdsMaps[jointId].at<float>(x, y);
float score = heatMapData[index];
allPeaks[jointId].reserve(maxNumPeople);
if (score > detectionThreshold) {
allPeaks[jointId].emplace_back(Peak{cv::Point2f(static_cast<float>(x), static_cast<float>(y)), score, tag});
}
}
}
std::vector<Pose> matchByTag(std::vector<std::vector<Peak>>& allPeaks,
size_t maxNumPeople,
size_t numJoints,
float tagThreshold) {
size_t jointOrder[]{0, 1, 2, 3, 4, 5, 6, 11, 12, 7, 8, 9, 10, 13, 14, 15, 16};
std::vector<Pose> allPoses;
for (size_t jointId : jointOrder) {
std::vector<Peak>& jointPeaks = allPeaks[jointId];
std::vector<float> tags;
for (auto& peak : jointPeaks) {
tags.push_back(peak.tag);
}
if (allPoses.empty()) {
for (size_t personId = 0; personId < jointPeaks.size(); personId++) {
Peak peak = jointPeaks[personId];
Pose pose = Pose(numJoints);
pose.add(jointId, peak);
allPoses.push_back(pose);
}
continue;
}
if (jointPeaks.empty() || (allPoses.size() == maxNumPeople)) {
continue;
}
std::vector<float> posesTags;
std::vector<cv::Point2f> posesCenters;
for (auto& pose : allPoses) {
posesTags.push_back(pose.getPoseTag());
posesCenters.push_back(pose.getPoseCenter());
}
size_t numAdded = tags.size();
size_t numGrouped = posesTags.size();
cv::Mat tagsDiff(numAdded, numGrouped, CV_32F);
cv::Mat matchingCost(numAdded, numGrouped, CV_32F);
std::vector<float> dists(numAdded);
for (size_t j = 0; j < numGrouped; j++) {
float minDist = std::numeric_limits<float>::max();
// Compute euclidean distance (in spatial space) between the pose center and all joints.
const cv::Point2f center = posesCenters.at(j);
for (size_t i = 0; i < numAdded; i++) {
cv::Point2f v = jointPeaks.at(i).keypoint - center;
float dist = std::sqrt(v.x * v.x + v.y * v.y);
dists[i] = dist;
minDist = std::min(dist, minDist);
}
// Compute semantic distance (in embedding space) between the pose tag and all joints
// and corresponding matching costs.
auto poseTag = posesTags[j];
for (size_t i = 0; i < numAdded; i++) {
float diff = static_cast<float>(cv::norm(tags[i] - poseTag));
tagsDiff.at<float>(i, j) = diff;
if (diff < tagThreshold) {
diff *= dists[i] / (minDist + 1e-10f);
}
matchingCost.at<float>(i, j) = std::round(diff) * 100 - jointPeaks[i].score;
}
}
if (numAdded > numGrouped) {
cv::copyMakeBorder(matchingCost,
matchingCost,
0,
0,
0,
numAdded - numGrouped,
cv::BORDER_CONSTANT,
10000000);
}
// Get pairs
auto res = KuhnMunkres().Solve(matchingCost);
for (size_t row = 0; row < res.size(); row++) {
size_t col = res[row];
if (row < numAdded && col < numGrouped && tagsDiff.at<float>(row, col) < tagThreshold) {
allPoses[col].add(jointId, jointPeaks[row]);
} else {
Pose pose = Pose(numJoints);
pose.add(jointId, jointPeaks[row]);
allPoses.push_back(pose);
}
}
}
return allPoses;
}
namespace {
cv::Point2f adjustLocation(const int x, const int y, const cv::Mat& heatMap) {
cv::Point2f delta(0.f, 0.f);
int width = heatMap.cols;
int height = heatMap.rows;
if ((1 < x) && (x < width - 1) && (1 < y) && (y < height - 1)) {
auto diffX = heatMap.at<float>(y, x + 1) - heatMap.at<float>(y, x - 1);
auto diffY = heatMap.at<float>(y + 1, x) - heatMap.at<float>(y - 1, x);
delta.x = diffX > 0 ? 0.25f : -0.25f;
delta.y = diffY > 0 ? 0.25f : -0.25f;
}
return delta;
}
} // namespace
void adjustAndRefine(std::vector<Pose>& allPoses,
const std::vector<cv::Mat>& heatMaps,
const std::vector<cv::Mat>& aembdsMaps,
int poseId,
const float delta) {
Pose& pose = allPoses[poseId];
float poseTag = pose.getPoseTag();
for (size_t jointId = 0; jointId < pose.size(); jointId++) {
Peak& peak = pose.getPeak(jointId);
const cv::Mat& heatMap = heatMaps[jointId];
const cv::Mat& aembds = aembdsMaps[jointId];
if (peak.score > 0) {
// Adjust
int x = static_cast<int>(peak.keypoint.x);
int y = static_cast<int>(peak.keypoint.y);
peak.keypoint += adjustLocation(x, y, heatMap);
if (delta) {
peak.keypoint.x += delta;
peak.keypoint.y += delta;
}
} else {
// Refine
// Get position with the closest tag value to the pose tag
cv::Mat diff = cv::abs(aembds - poseTag);
diff.convertTo(diff, CV_32S, 1.0, 0.0);
diff.convertTo(diff, CV_32F);
diff -= heatMap;
double min;
cv::Point2i minLoc;
cv::minMaxLoc(diff, &min, 0, &minLoc);
int x = minLoc.x;
int y = minLoc.y;
float val = heatMap.at<float>(y, x);
if (val > 0) {
peak.keypoint.x = static_cast<float>(x);
peak.keypoint.y = static_cast<float>(y);
peak.keypoint += adjustLocation(x, y, heatMap);
// Peak score is assigned directly, so it does not affect the pose score.
peak.score = val;
}
}
}
}
|
