Visual Servoing Platform  version 3.2.0 under development (2019-01-22)
vpDetectorDNN.cpp
1 /****************************************************************************
2  *
3  * ViSP, open source Visual Servoing Platform software.
4  * Copyright (C) 2005 - 2019 by Inria. All rights reserved.
5  *
6  * This software is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License as published by
8  * the Free Software Foundation; either version 2 of the License, or
9  * (at your option) any later version.
10  * See the file LICENSE.txt at the root directory of this source
11  * distribution for additional information about the GNU GPL.
12  *
13  * For using ViSP with software that can not be combined with the GNU
14  * GPL, please contact Inria about acquiring a ViSP Professional
15  * Edition License.
16  *
17  * See http://visp.inria.fr for more information.
18  *
19  * This software was developed at:
20  * Inria Rennes - Bretagne Atlantique
21  * Campus Universitaire de Beaulieu
22  * 35042 Rennes Cedex
23  * France
24  *
25  * If you have questions regarding the use of this file, please contact
26  * Inria at visp@inria.fr
27  *
28  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
29  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
30  *
31  * Description:
32  * DNN object detection using OpenCV DNN module.
33  *
34  *****************************************************************************/
35 #include <visp3/core/vpConfig.h>
36 
37 #if (VISP_HAVE_OPENCV_VERSION >= 0x030403)
38 #include <visp3/detection/vpDetectorDNN.h>
39 #include <visp3/core/vpImageConvert.h>
40 
41 vpDetectorDNN::vpDetectorDNN() : m_blob(), m_boxes(), m_classIds(), m_confidences(),
42  m_confidenceThreshold(0.5), m_I_color(), m_img(), m_inputSize(300,300), m_mean(127.5, 127.5, 127.5),
43  m_net(), m_nmsThreshold(0.4f), m_outNames(), m_outs(), m_scaleFactor(2.0/255.0), m_swapRB(true) {}
44 
46 
56  vpImageConvert::convert(I, m_I_color);
57 
58  std::vector<vpRect> boundingBoxes;
59  return detect(m_I_color, boundingBoxes);
60 }
61 
69 bool vpDetectorDNN::detect(const vpImage<vpRGBa> &I, std::vector<vpRect> &boundingBoxes) {
70  vpImageConvert::convert(I, m_img);
71 
72  cv::Size inputSize(m_inputSize.width > 0 ? m_inputSize.width : m_img.cols,
73  m_inputSize.height > 0 ? m_inputSize.height : m_img.rows);
74  cv::dnn::blobFromImage(m_img, m_blob, m_scaleFactor, inputSize, m_mean, m_swapRB, false);
75 
76  m_net.setInput(m_blob);
77  m_net.forward(m_outs, m_outNames);
78 
79  postProcess();
80 
81  boundingBoxes.resize(m_boxesNMS.size());
82  for (size_t i = 0; i < m_boxesNMS.size(); i++) {
83  cv::Rect box = m_boxesNMS[i];
84  boundingBoxes[i] = vpRect(box.x, box.y, box.width, box.height);
85  }
86 
87  m_nb_objects = boundingBoxes.size();
88  m_polygon.resize(boundingBoxes.size());
89  m_message.resize(boundingBoxes.size());
90  for (size_t i = 0; i < boundingBoxes.size(); i++) {
91  std::vector<vpImagePoint> polygon;
92 
93  double x = boundingBoxes[i].getLeft();
94  double y = boundingBoxes[i].getTop();
95  double w = boundingBoxes[i].getWidth();
96  double h = boundingBoxes[i].getHeight();
97 
98  polygon.push_back(vpImagePoint(y, x));
99  polygon.push_back(vpImagePoint(y + h, x));
100  polygon.push_back(vpImagePoint(y + h, x + w));
101  polygon.push_back(vpImagePoint(y, x + w));
102 
103  m_polygon[i] = polygon;
104 
105  std::ostringstream oss;
106  oss << m_classIds[i] << " ; " << m_confidences[i] << " ; " << m_boxes[i];
107  m_message[i] = oss.str();
108  }
109 
110  return !boundingBoxes.empty();
111 }
112 
118 std::vector<vpRect> vpDetectorDNN::getDetectionBBs(bool afterNMS) const {
119  std::vector<vpRect> bbs;
120  if (afterNMS) {
121  bbs.reserve(m_boxesNMS.size());
122  for (size_t i = 0; i < m_boxesNMS.size(); i++) {
123  cv::Rect box = m_boxes[i];
124  bbs.push_back(vpRect(box.x, box.y, box.width, box.height));
125  }
126  } else {
127  bbs.reserve(m_boxes.size());
128  for (size_t i = 0; i < m_boxes.size(); i++) {
129  cv::Rect box = m_boxes[i];
130  bbs.push_back(vpRect(box.x, box.y, box.width, box.height));
131  }
132  }
133 
134  return bbs;
135 }
136 
142 std::vector<int> vpDetectorDNN::getDetectionClassIds(bool afterNMS) const {
143  if (afterNMS) {
144  std::vector<int> classIds;
145  for (size_t i = 0; i < m_indices.size(); i++) {
146  int idx = m_indices[i];
147  classIds.push_back(m_classIds[idx]);
148  }
149  return classIds;
150  }
151 
152  return m_classIds;
153 }
154 
158 std::vector<float> vpDetectorDNN::getDetectionConfidence(bool afterNMS) const {
159  if (afterNMS) {
160  std::vector<float> confidences;
161  for (size_t i = 0; i < m_indices.size(); i++) {
162  int idx = m_indices[i];
163  confidences.push_back(m_confidences[idx]);
164  }
165  return confidences;
166  }
167 
168  return m_confidences;
169 }
170 
171 #if (VISP_HAVE_OPENCV_VERSION == 0x030403)
172 std::vector<cv::String> vpDetectorDNN::getOutputsNames() {
173  static std::vector<cv::String> names;
174  if (names.empty()) {
175  std::vector<int> outLayers = m_net.getUnconnectedOutLayers();
176  std::vector<cv::String> layersNames = m_net.getLayerNames();
177  names.resize(outLayers.size());
178  for (size_t i = 0; i < outLayers.size(); ++i)
179  names[i] = layersNames[outLayers[i] - 1];
180  }
181  return names;
182 }
183 #endif
184 
185 void vpDetectorDNN::postProcess() {
186  //Direct copy from object_detection.cpp OpenCV sample
187  static std::vector<int> outLayers = m_net.getUnconnectedOutLayers();
188  static std::string outLayerType = m_net.getLayer(outLayers[0])->type;
189 
190  m_classIds.clear();
191  m_confidences.clear();
192  m_boxes.clear();
193  if (m_net.getLayer(0)->outputNameToIndex("im_info") != -1) // Faster-RCNN or R-FCN
194  {
195  // Network produces output blob with a shape 1x1xNx7 where N is a number of
196  // detections and an every detection is a vector of values
197  // [batchId, classId, confidence, left, top, right, bottom]
198  CV_Assert(m_outs.size() == 1);
199  float* data = (float*)m_outs[0].data;
200  for (size_t i = 0; i < m_outs[0].total(); i += 7)
201  {
202  float confidence = data[i + 2];
203  if (confidence > m_confidenceThreshold)
204  {
205  int left = (int)data[i + 3];
206  int top = (int)data[i + 4];
207  int right = (int)data[i + 5];
208  int bottom = (int)data[i + 6];
209  int width = right - left + 1;
210  int height = bottom - top + 1;
211  m_classIds.push_back((int)(data[i + 1]) - 1); // Skip 0th background class id.
212  m_boxes.push_back(cv::Rect(left, top, width, height));
213  m_confidences.push_back(confidence);
214  }
215  }
216  }
217  else if (outLayerType == "DetectionOutput")
218  {
219  // Network produces output blob with a shape 1x1xNx7 where N is a number of
220  // detections and an every detection is a vector of values
221  // [batchId, classId, confidence, left, top, right, bottom]
222  CV_Assert(m_outs.size() == 1);
223  float* data = (float*)m_outs[0].data;
224  for (size_t i = 0; i < m_outs[0].total(); i += 7)
225  {
226  float confidence = data[i + 2];
227  if (confidence > m_confidenceThreshold)
228  {
229  int left = (int)(data[i + 3] * m_img.cols);
230  int top = (int)(data[i + 4] * m_img.rows);
231  int right = (int)(data[i + 5] * m_img.cols);
232  int bottom = (int)(data[i + 6] * m_img.rows);
233  int width = right - left + 1;
234  int height = bottom - top + 1;
235  m_classIds.push_back((int)(data[i + 1]) - 1); // Skip 0th background class id.
236  m_boxes.push_back(cv::Rect(left, top, width, height));
237  m_confidences.push_back(confidence);
238  }
239  }
240  }
241  else if (outLayerType == "Region")
242  {
243  for (size_t i = 0; i < m_outs.size(); ++i)
244  {
245  // Network produces output blob with a shape NxC where N is a number of
246  // detected objects and C is a number of classes + 4 where the first 4
247  // numbers are [center_x, center_y, width, height]
248  float* data = (float*)m_outs[i].data;
249  for (int j = 0; j < m_outs[i].rows; ++j, data += m_outs[i].cols)
250  {
251  cv::Mat scores = m_outs[i].row(j).colRange(5, m_outs[i].cols);
252  cv::Point classIdPoint;
253  double confidence;
254  cv::minMaxLoc(scores, 0, &confidence, 0, &classIdPoint);
255  if (confidence > m_confidenceThreshold)
256  {
257  int centerX = (int)(data[0] * m_img.cols);
258  int centerY = (int)(data[1] * m_img.rows);
259  int width = (int)(data[2] * m_img.cols);
260  int height = (int)(data[3] * m_img.rows);
261  int left = centerX - width / 2;
262  int top = centerY - height / 2;
263 
264  m_classIds.push_back(classIdPoint.x);
265  m_confidences.push_back((float)confidence);
266  m_boxes.push_back(cv::Rect(left, top, width, height));
267  }
268  }
269  }
270  }
271  else
272  CV_Error(cv::Error::StsNotImplemented, "Unknown output layer type: " + outLayerType);
273 
274  cv::dnn::NMSBoxes(m_boxes, m_confidences, m_confidenceThreshold, m_nmsThreshold, m_indices);
275  m_boxesNMS.resize(m_indices.size());
276  for (size_t i = 0; i < m_indices.size(); ++i)
277  {
278  int idx = m_indices[i];
279  m_boxesNMS[i] = m_boxes[idx];
280  }
281 }
282 
301 void vpDetectorDNN::readNet(const std::string &model, const std::string &config, const std::string &framework) {
302  m_net = cv::dnn::readNet(model, config, framework);
303 #if (VISP_HAVE_OPENCV_VERSION == 0x030403)
304  m_outNames = getOutputsNames();
305 #else
306  m_outNames = m_net.getUnconnectedOutLayersNames();
307 #endif
308 }
309 
315 void vpDetectorDNN::setConfidenceThreshold(float confThreshold) {
316  m_confidenceThreshold = confThreshold;
317 }
318 
325 void vpDetectorDNN::setInputSize(int width, int height) {
326  m_inputSize.width = width;
327  m_inputSize.height = height;
328 }
329 
337 void vpDetectorDNN::setMean(double meanR, double meanG, double meanB) {
338  m_mean = cv::Scalar(meanR, meanG, meanB);
339 }
340 
347 void vpDetectorDNN::setNMSThreshold(float nmsThreshold) {
348  m_nmsThreshold = nmsThreshold;
349 }
350 
358  m_net.setPreferableBackend(backendId);
359 }
360 
368  m_net.setPreferableTarget(targetId);
369 }
370 
374 void vpDetectorDNN::setScaleFactor(double scaleFactor) {
375  m_scaleFactor = scaleFactor;
376 }
377 
383 void vpDetectorDNN::setSwapRB(bool swapRB) {
384  m_swapRB = swapRB;
385 }
386 
387 #elif !defined(VISP_BUILD_SHARED_LIBS)
388 // Work arround to avoid warning: libvisp_core.a(vpDetectorDNN.cpp.o) has no
389 // symbols
390 void dummy_vpDetectorDNN(){};
391 #endif
void setMean(double meanR, double meanG, double meanB)
static void convert(const vpImage< unsigned char > &src, vpImage< vpRGBa > &dest)
void setPreferableTarget(int targetId)
std::vector< int > getDetectionClassIds(bool afterNMS=true) const
void setPreferableBackend(int backendId)
void readNet(const std::string &model, const std::string &config="", const std::string &framework="")
std::vector< vpRect > getDetectionBBs(bool afterNMS=true) const
size_t m_nb_objects
Number of detected objects.
void setSwapRB(bool swapRB)
virtual ~vpDetectorDNN()
void setScaleFactor(double scaleFactor)
void setNMSThreshold(float nmsThreshold)
void setInputSize(int width, int height)
void setConfidenceThreshold(float confThreshold)
std::vector< std::vector< vpImagePoint > > m_polygon
Defines a rectangle in the plane.
Definition: vpRect.h:78
Class that defines a 2D point in an image. This class is useful for image processing and stores only ...
Definition: vpImagePoint.h:88
std::vector< std::string > m_message
Message attached to each object.
virtual bool detect(const vpImage< unsigned char > &I)
std::vector< float > getDetectionConfidence(bool afterNMS=true) const