| 1 | /**** |
| 2 | Copyright (c) 2020 Adrian I. Lam |
| 3 | |
| 4 | Permission is hereby granted, free of charge, to any person obtaining a copy |
| 5 | of this software and associated documentation files (the "Software"), to deal |
| 6 | in the Software without restriction, including without limitation the rights |
| 7 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
| 8 | copies of the Software, and to permit persons to whom the Software is |
| 9 | furnished to do so, subject to the following conditions: |
| 10 | |
| 11 | The above copyright notice and this permission notice shall be included in all |
| 12 | copies or substantial portions of the Software. |
| 13 | |
| 14 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 15 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 16 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 17 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 18 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 19 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 20 | SOFTWARE. |
| 21 | ****/ |
| 22 | |
| 23 | #include <stdexcept> |
| 24 | #include <fstream> |
| 25 | #include <string> |
| 26 | #include <sstream> |
| 27 | #include <cmath> |
| 28 | |
| 29 | #include <opencv2/opencv.hpp> |
| 30 | |
| 31 | #include <dlib/opencv.h> |
| 32 | #include <dlib/image_processing/frontal_face_detector.h> |
| 33 | #include <dlib/image_processing.h> |
| 34 | #include <dlib/image_processing/render_face_detections.h> |
| 35 | |
| 36 | #include "facial_landmark_detector.h" |
| 37 | #include "math_utils.h" |
| 38 | |
| 39 | |
| 40 | static void filterPush(std::deque<double>& buf, double newval, |
| 41 | std::size_t numTaps) |
| 42 | { |
| 43 | buf.push_back(newval); |
| 44 | while (buf.size() > numTaps) |
| 45 | { |
| 46 | buf.pop_front(); |
| 47 | } |
| 48 | } |
| 49 | |
| 50 | FacialLandmarkDetector::FacialLandmarkDetector(std::string cfgPath) |
| 51 | : m_stop(false) |
| 52 | { |
| 53 | parseConfig(cfgPath); |
| 54 | |
| 55 | if (!webcam.open(m_cfg.cvVideoCaptureId)) |
| 56 | { |
| 57 | throw std::runtime_error("Unable to open webcam"); |
| 58 | } |
| 59 | |
| 60 | detector = dlib::get_frontal_face_detector(); |
| 61 | dlib::deserialize(m_cfg.predictorPath) >> predictor; |
| 62 | } |
| 63 | |
| 64 | FacialLandmarkDetector::Params FacialLandmarkDetector::getParams(void) const |
| 65 | { |
| 66 | Params params; |
| 67 | |
| 68 | params.faceXAngle = avg(m_faceXAngle); |
| 69 | params.faceYAngle = avg(m_faceYAngle) + m_cfg.faceYAngleCorrection; |
| 70 | // + 10 correct for angle between computer monitor and webcam |
| 71 | params.faceZAngle = avg(m_faceZAngle); |
| 72 | params.mouthOpenness = avg(m_mouthOpenness); |
| 73 | params.mouthForm = avg(m_mouthForm); |
| 74 | |
| 75 | double leftEye = avg(m_leftEyeOpenness, 1); |
| 76 | double rightEye = avg(m_rightEyeOpenness, 1); |
| 77 | // Just combine the two to get better synchronized blinks |
| 78 | // This effectively disables winks, so if we want to |
| 79 | // support winks in the future (see below) we will need |
| 80 | // a better way to handle this out-of-sync blinks. |
| 81 | double bothEyes = (leftEye + rightEye) / 2; |
| 82 | leftEye = bothEyes; |
| 83 | rightEye = bothEyes; |
| 84 | // Detect winks and make them look better |
| 85 | // Commenting out - winks are difficult to be detected by the |
| 86 | // dlib data set anyway... maybe in the future we can |
| 87 | // add a runtime option to enable/disable... |
| 88 | /*if (right == 0 && left > 0.2) |
| 89 | { |
| 90 | left = 1; |
| 91 | } |
| 92 | else if (left == 0 && right > 0.2) |
| 93 | { |
| 94 | right = 1; |
| 95 | } |
| 96 | */ |
| 97 | params.leftEyeOpenness = leftEye; |
| 98 | params.rightEyeOpenness = rightEye; |
| 99 | |
| 100 | if (leftEye <= m_cfg.eyeSmileEyeOpenThreshold && |
| 101 | rightEye <= m_cfg.eyeSmileEyeOpenThreshold && |
| 102 | params.mouthForm > m_cfg.eyeSmileMouthFormThreshold && |
| 103 | params.mouthOpenness > m_cfg.eyeSmileMouthOpenThreshold) |
| 104 | { |
| 105 | params.leftEyeSmile = 1; |
| 106 | params.rightEyeSmile = 1; |
| 107 | } |
| 108 | else |
| 109 | { |
| 110 | params.leftEyeSmile = 0; |
| 111 | params.rightEyeSmile = 0; |
| 112 | } |
| 113 | |
| 114 | params.autoBlink = m_cfg.autoBlink; |
| 115 | params.autoBreath = m_cfg.autoBreath; |
| 116 | params.randomMotion = m_cfg.randomMotion; |
| 117 | |
| 118 | return params; |
| 119 | } |
| 120 | |
| 121 | void FacialLandmarkDetector::stop(void) |
| 122 | { |
| 123 | m_stop = true; |
| 124 | } |
| 125 | |
| 126 | void FacialLandmarkDetector::mainLoop(void) |
| 127 | { |
| 128 | while (!m_stop) |
| 129 | { |
| 130 | cv::Mat frame; |
| 131 | if (!webcam.read(frame)) |
| 132 | { |
| 133 | throw std::runtime_error("Unable to read from webcam"); |
| 134 | } |
| 135 | cv::Mat flipped; |
| 136 | if (m_cfg.lateralInversion) |
| 137 | { |
| 138 | cv::flip(frame, flipped, 1); |
| 139 | } |
| 140 | else |
| 141 | { |
| 142 | flipped = frame; |
| 143 | } |
| 144 | dlib::cv_image<dlib::bgr_pixel> cimg(flipped); |
| 145 | |
| 146 | if (m_cfg.showWebcamVideo) |
| 147 | { |
| 148 | win.set_image(cimg); |
| 149 | } |
| 150 | |
| 151 | std::vector<dlib::rectangle> faces = detector(cimg); |
| 152 | |
| 153 | if (faces.size() > 0) |
| 154 | { |
| 155 | dlib::rectangle face = faces[0]; |
| 156 | dlib::full_object_detection shape = predictor(cimg, face); |
| 157 | |
| 158 | /* The coordinates seem to be rather noisy in general. |
| 159 | * We will push everything through some moving average filters |
| 160 | * to reduce noise. The number of taps is determined empirically |
| 161 | * until we get something good. |
| 162 | * An alternative method would be to get some better dataset |
| 163 | * for dlib - perhaps even to train on a custom data set just for the user. |
| 164 | */ |
| 165 | |
| 166 | // Face rotation: X direction (left-right) |
| 167 | double faceXRot = calcFaceXAngle(shape); |
| 168 | filterPush(m_faceXAngle, faceXRot, m_cfg.faceXAngleNumTaps); |
| 169 | |
| 170 | // Mouth form (smile / laugh) detection |
| 171 | double mouthForm = calcMouthForm(shape); |
| 172 | filterPush(m_mouthForm, mouthForm, m_cfg.mouthFormNumTaps); |
| 173 | |
| 174 | // Face rotation: Y direction (up-down) |
| 175 | double faceYRot = calcFaceYAngle(shape, faceXRot, mouthForm); |
| 176 | filterPush(m_faceYAngle, faceYRot, m_cfg.faceYAngleNumTaps); |
| 177 | |
| 178 | // Face rotation: Z direction (head tilt) |
| 179 | double faceZRot = calcFaceZAngle(shape); |
| 180 | filterPush(m_faceZAngle, faceZRot, m_cfg.faceZAngleNumTaps); |
| 181 | |
| 182 | // Mouth openness |
| 183 | double mouthOpen = calcMouthOpenness(shape, mouthForm); |
| 184 | filterPush(m_mouthOpenness, mouthOpen, m_cfg.mouthOpenNumTaps); |
| 185 | |
| 186 | // Eye openness |
| 187 | double eyeLeftOpen = calcEyeOpenness(LEFT, shape, faceYRot); |
| 188 | filterPush(m_leftEyeOpenness, eyeLeftOpen, m_cfg.leftEyeOpenNumTaps); |
| 189 | double eyeRightOpen = calcEyeOpenness(RIGHT, shape, faceYRot); |
| 190 | filterPush(m_rightEyeOpenness, eyeRightOpen, m_cfg.rightEyeOpenNumTaps); |
| 191 | |
| 192 | // TODO eyebrows? |
| 193 | |
| 194 | if (m_cfg.showWebcamVideo && m_cfg.renderLandmarksOnVideo) |
| 195 | { |
| 196 | win.clear_overlay(); |
| 197 | win.add_overlay(dlib::render_face_detections(shape)); |
| 198 | } |
| 199 | } |
| 200 | else |
| 201 | { |
| 202 | if (m_cfg.showWebcamVideo && m_cfg.renderLandmarksOnVideo) |
| 203 | { |
| 204 | win.clear_overlay(); |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | cv::waitKey(m_cfg.cvWaitKeyMs); |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | double FacialLandmarkDetector::calcEyeAspectRatio( |
| 213 | dlib::point& p1, dlib::point& p2, |
| 214 | dlib::point& p3, dlib::point& p4, |
| 215 | dlib::point& p5, dlib::point& p6) const |
| 216 | { |
| 217 | double eyeWidth = dist(p1, p4); |
| 218 | double eyeHeight1 = dist(p2, p6); |
| 219 | double eyeHeight2 = dist(p3, p5); |
| 220 | |
| 221 | return (eyeHeight1 + eyeHeight2) / (2 * eyeWidth); |
| 222 | } |
| 223 | |
| 224 | double FacialLandmarkDetector::calcEyeOpenness( |
| 225 | LeftRight eye, |
| 226 | dlib::full_object_detection& shape, |
| 227 | double faceYAngle) const |
| 228 | { |
| 229 | double eyeAspectRatio; |
| 230 | if (eye == LEFT) |
| 231 | { |
| 232 | eyeAspectRatio = calcEyeAspectRatio(shape.part(42), shape.part(43), shape.part(44), |
| 233 | shape.part(45), shape.part(46), shape.part(47)); |
| 234 | } |
| 235 | else |
| 236 | { |
| 237 | eyeAspectRatio = calcEyeAspectRatio(shape.part(36), shape.part(37), shape.part(38), |
| 238 | shape.part(39), shape.part(40), shape.part(41)); |
| 239 | } |
| 240 | |
| 241 | // Apply correction due to faceYAngle |
| 242 | double corrEyeAspRat = eyeAspectRatio / std::cos(degToRad(faceYAngle)); |
| 243 | |
| 244 | return linearScale01(corrEyeAspRat, m_cfg.eyeClosedThreshold, m_cfg.eyeOpenThreshold); |
| 245 | } |
| 246 | |
| 247 | |
| 248 | |
| 249 | double FacialLandmarkDetector::calcMouthForm(dlib::full_object_detection& shape) const |
| 250 | { |
| 251 | /* Mouth form parameter: 0 for normal mouth, 1 for fully smiling / laughing. |
| 252 | * Compare distance between the two corners of the mouth |
| 253 | * to the distance between the two eyes. |
| 254 | */ |
| 255 | |
| 256 | /* An alternative (my initial attempt) was to compare the corners of |
| 257 | * the mouth to the top of the upper lip - they almost lie on a |
| 258 | * straight line when smiling / laughing. But that is only true |
| 259 | * when facing straight at the camera. When looking up / down, |
| 260 | * the angle changes. So here we'll use the distance approach instead. |
| 261 | */ |
| 262 | |
| 263 | auto eye1 = centroid(shape.part(36), shape.part(37), shape.part(38), |
| 264 | shape.part(39), shape.part(40), shape.part(41)); |
| 265 | auto eye2 = centroid(shape.part(42), shape.part(43), shape.part(44), |
| 266 | shape.part(45), shape.part(46), shape.part(47)); |
| 267 | double distEyes = dist(eye1, eye2); |
| 268 | double distMouth = dist(shape.part(48), shape.part(54)); |
| 269 | |
| 270 | double form = linearScale01(distMouth / distEyes, |
| 271 | m_cfg.mouthNormalThreshold, |
| 272 | m_cfg.mouthSmileThreshold); |
| 273 | |
| 274 | return form; |
| 275 | } |
| 276 | |
| 277 | double FacialLandmarkDetector::calcMouthOpenness( |
| 278 | dlib::full_object_detection& shape, |
| 279 | double mouthForm) const |
| 280 | { |
| 281 | // Use points for the bottom of the upper lip, and top of the lower lip |
| 282 | // We have 3 pairs of points available, which give the mouth height |
| 283 | // on the left, in the middle, and on the right, resp. |
| 284 | // First let's try to use an average of all three. |
| 285 | double heightLeft = dist(shape.part(63), shape.part(65)); |
| 286 | double heightMiddle = dist(shape.part(62), shape.part(66)); |
| 287 | double heightRight = dist(shape.part(61), shape.part(67)); |
| 288 | |
| 289 | double avgHeight = (heightLeft + heightMiddle + heightRight) / 3; |
| 290 | |
| 291 | // Now, normalize it with the width of the mouth. |
| 292 | double width = dist(shape.part(60), shape.part(64)); |
| 293 | |
| 294 | double normalized = avgHeight / width; |
| 295 | |
| 296 | double scaled = linearScale01(normalized, |
| 297 | m_cfg.mouthClosedThreshold, |
| 298 | m_cfg.mouthOpenThreshold, |
| 299 | true, false); |
| 300 | |
| 301 | // Apply correction according to mouthForm |
| 302 | // Notice that when you smile / laugh, width is increased |
| 303 | scaled *= (1 + m_cfg.mouthOpenLaughCorrection * mouthForm); |
| 304 | |
| 305 | return scaled; |
| 306 | } |
| 307 | |
| 308 | double FacialLandmarkDetector::calcFaceXAngle(dlib::full_object_detection& shape) const |
| 309 | { |
| 310 | // This function will be easier to understand if you refer to the |
| 311 | // diagram in faceXAngle.png |
| 312 | |
| 313 | // Construct the y-axis using (1) average of four points on the nose and |
| 314 | // (2) average of four points on the upper lip. |
| 315 | |
| 316 | auto y0 = centroid(shape.part(27), shape.part(28), shape.part(29), |
| 317 | shape.part(30)); |
| 318 | auto y1 = centroid(shape.part(50), shape.part(51), shape.part(52), |
| 319 | shape.part(62)); |
| 320 | |
| 321 | // Now drop a perpedicular from the left and right edges of the face, |
| 322 | // and calculate the ratio between the lengths of these perpendiculars |
| 323 | |
| 324 | auto left = centroid(shape.part(14), shape.part(15), shape.part(16)); |
| 325 | auto right = centroid(shape.part(0), shape.part(1), shape.part(2)); |
| 326 | |
| 327 | // Constructing a perpendicular: |
| 328 | // Join the left/right point and the upper lip. The included angle |
| 329 | // can now be determined using cosine rule. |
| 330 | // Then sine of this angle is the perpendicular divided by the newly |
| 331 | // created line. |
| 332 | double opp = dist(right, y0); |
| 333 | double adj1 = dist(y0, y1); |
| 334 | double adj2 = dist(y1, right); |
| 335 | double angle = solveCosineRuleAngle(opp, adj1, adj2); |
| 336 | double perpRight = adj2 * std::sin(angle); |
| 337 | |
| 338 | opp = dist(left, y0); |
| 339 | adj2 = dist(y1, left); |
| 340 | angle = solveCosineRuleAngle(opp, adj1, adj2); |
| 341 | double perpLeft = adj2 * std::sin(angle); |
| 342 | |
| 343 | // Model the head as a sphere and look from above. |
| 344 | double theta = std::asin((perpRight - perpLeft) / (perpRight + perpLeft)); |
| 345 | |
| 346 | theta = radToDeg(theta); |
| 347 | if (theta < -30) theta = -30; |
| 348 | if (theta > 30) theta = 30; |
| 349 | return theta; |
| 350 | } |
| 351 | |
| 352 | double FacialLandmarkDetector::calcFaceYAngle(dlib::full_object_detection& shape, double faceXAngle, double mouthForm) const |
| 353 | { |
| 354 | // Use the nose |
| 355 | // angle between the two left/right points and the tip |
| 356 | double c = dist(shape.part(31), shape.part(35)); |
| 357 | double a = dist(shape.part(30), shape.part(31)); |
| 358 | double b = dist(shape.part(30), shape.part(35)); |
| 359 | |
| 360 | double angle = solveCosineRuleAngle(c, a, b); |
| 361 | |
| 362 | // This probably varies a lot from person to person... |
| 363 | |
| 364 | // Best is probably to work out some trigonometry again, |
| 365 | // but just linear interpolation seems to work ok... |
| 366 | |
| 367 | // Correct for X rotation |
| 368 | double corrAngle = angle * (1 + (std::abs(faceXAngle) / 30 |
| 369 | * m_cfg.faceYAngleXRotCorrection)); |
| 370 | |
| 371 | // Correct for smiles / laughs - this increases the angle |
| 372 | corrAngle *= (1 - mouthForm * m_cfg.faceYAngleSmileCorrection); |
| 373 | |
| 374 | if (corrAngle >= m_cfg.faceYAngleZeroValue) |
| 375 | { |
| 376 | return -30 * linearScale01(corrAngle, |
| 377 | m_cfg.faceYAngleZeroValue, |
| 378 | m_cfg.faceYAngleDownThreshold, |
| 379 | false, false); |
| 380 | } |
| 381 | else |
| 382 | { |
| 383 | return 30 * (1 - linearScale01(corrAngle, |
| 384 | m_cfg.faceYAngleUpThreshold, |
| 385 | m_cfg.faceYAngleZeroValue, |
| 386 | false, false)); |
| 387 | } |
| 388 | } |
| 389 | |
| 390 | double FacialLandmarkDetector::calcFaceZAngle(dlib::full_object_detection& shape) const |
| 391 | { |
| 392 | // Use average of eyes and nose |
| 393 | |
| 394 | auto eyeRight = centroid(shape.part(36), shape.part(37), shape.part(38), |
| 395 | shape.part(39), shape.part(40), shape.part(41)); |
| 396 | auto eyeLeft = centroid(shape.part(42), shape.part(43), shape.part(44), |
| 397 | shape.part(45), shape.part(46), shape.part(47)); |
| 398 | |
| 399 | auto noseLeft = shape.part(35); |
| 400 | auto noseRight = shape.part(31); |
| 401 | |
| 402 | double eyeYDiff = eyeRight.y() - eyeLeft.y(); |
| 403 | double eyeXDiff = eyeRight.x() - eyeLeft.x(); |
| 404 | |
| 405 | double angle1 = std::atan(eyeYDiff / eyeXDiff); |
| 406 | |
| 407 | double noseYDiff = noseRight.y() - noseLeft.y(); |
| 408 | double noseXDiff = noseRight.x() - noseLeft.x(); |
| 409 | |
| 410 | double angle2 = std::atan(noseYDiff / noseXDiff); |
| 411 | |
| 412 | return radToDeg((angle1 + angle2) / 2); |
| 413 | } |
| 414 | |
| 415 | void FacialLandmarkDetector::parseConfig(std::string cfgPath) |
| 416 | { |
| 417 | populateDefaultConfig(); |
| 418 | if (cfgPath != "") |
| 419 | { |
| 420 | std::ifstream file(cfgPath); |
| 421 | |
| 422 | if (!file) |
| 423 | { |
| 424 | throw std::runtime_error("Failed to open config file"); |
| 425 | } |
| 426 | |
| 427 | std::string line; |
| 428 | unsigned int lineNum = 0; |
| 429 | |
| 430 | while (std::getline(file, line)) |
| 431 | { |
| 432 | lineNum++; |
| 433 | |
| 434 | if (line[0] == '#') |
| 435 | { |
| 436 | continue; |
| 437 | } |
| 438 | |
| 439 | std::istringstream ss(line); |
| 440 | std::string paramName; |
| 441 | if (ss >> paramName) |
| 442 | { |
| 443 | if (paramName == "cvVideoCaptureId") |
| 444 | { |
| 445 | if (!(ss >> m_cfg.cvVideoCaptureId)) |
| 446 | { |
| 447 | throwConfigError(paramName, "int", |
| 448 | line, lineNum); |
| 449 | } |
| 450 | } |
| 451 | else if (paramName == "predictorPath") |
| 452 | { |
| 453 | if (!(ss >> m_cfg.predictorPath)) |
| 454 | { |
| 455 | throwConfigError(paramName, "std::string", |
| 456 | line, lineNum); |
| 457 | } |
| 458 | } |
| 459 | else if (paramName == "faceYAngleCorrection") |
| 460 | { |
| 461 | if (!(ss >> m_cfg.faceYAngleCorrection)) |
| 462 | { |
| 463 | throwConfigError(paramName, "double", |
| 464 | line, lineNum); |
| 465 | } |
| 466 | } |
| 467 | else if (paramName == "eyeSmileEyeOpenThreshold") |
| 468 | { |
| 469 | if (!(ss >> m_cfg.eyeSmileEyeOpenThreshold)) |
| 470 | { |
| 471 | throwConfigError(paramName, "double", |
| 472 | line, lineNum); |
| 473 | } |
| 474 | } |
| 475 | else if (paramName == "eyeSmileMouthFormThreshold") |
| 476 | { |
| 477 | if (!(ss >> m_cfg.eyeSmileMouthFormThreshold)) |
| 478 | { |
| 479 | throwConfigError(paramName, "double", |
| 480 | line, lineNum); |
| 481 | } |
| 482 | } |
| 483 | else if (paramName == "eyeSmileMouthOpenThreshold") |
| 484 | { |
| 485 | if (!(ss >> m_cfg.eyeSmileMouthOpenThreshold)) |
| 486 | { |
| 487 | throwConfigError(paramName, "double", |
| 488 | line, lineNum); |
| 489 | } |
| 490 | } |
| 491 | else if (paramName == "showWebcamVideo") |
| 492 | { |
| 493 | if (!(ss >> m_cfg.showWebcamVideo)) |
| 494 | { |
| 495 | throwConfigError(paramName, "bool", |
| 496 | line, lineNum); |
| 497 | } |
| 498 | } |
| 499 | else if (paramName == "renderLandmarksOnVideo") |
| 500 | { |
| 501 | if (!(ss >> m_cfg.renderLandmarksOnVideo)) |
| 502 | { |
| 503 | throwConfigError(paramName, "bool", |
| 504 | line, lineNum); |
| 505 | } |
| 506 | } |
| 507 | else if (paramName == "lateralInversion") |
| 508 | { |
| 509 | if (!(ss >> m_cfg.lateralInversion)) |
| 510 | { |
| 511 | throwConfigError(paramName, "bool", |
| 512 | line, lineNum); |
| 513 | } |
| 514 | } |
| 515 | else if (paramName == "faceXAngleNumTaps") |
| 516 | { |
| 517 | if (!(ss >> m_cfg.faceXAngleNumTaps)) |
| 518 | { |
| 519 | throwConfigError(paramName, "std::size_t", |
| 520 | line, lineNum); |
| 521 | } |
| 522 | } |
| 523 | else if (paramName == "faceYAngleNumTaps") |
| 524 | { |
| 525 | if (!(ss >> m_cfg.faceYAngleNumTaps)) |
| 526 | { |
| 527 | throwConfigError(paramName, "std::size_t", |
| 528 | line, lineNum); |
| 529 | } |
| 530 | } |
| 531 | else if (paramName == "faceZAngleNumTaps") |
| 532 | { |
| 533 | if (!(ss >> m_cfg.faceZAngleNumTaps)) |
| 534 | { |
| 535 | throwConfigError(paramName, "std::size_t", |
| 536 | line, lineNum); |
| 537 | } |
| 538 | } |
| 539 | else if (paramName == "mouthFormNumTaps") |
| 540 | { |
| 541 | if (!(ss >> m_cfg.mouthFormNumTaps)) |
| 542 | { |
| 543 | throwConfigError(paramName, "std::size_t", |
| 544 | line, lineNum); |
| 545 | } |
| 546 | } |
| 547 | else if (paramName == "mouthOpenNumTaps") |
| 548 | { |
| 549 | if (!(ss >> m_cfg.mouthOpenNumTaps)) |
| 550 | { |
| 551 | throwConfigError(paramName, "std::size_t", |
| 552 | line, lineNum); |
| 553 | } |
| 554 | } |
| 555 | else if (paramName == "leftEyeOpenNumTaps") |
| 556 | { |
| 557 | if (!(ss >> m_cfg.leftEyeOpenNumTaps)) |
| 558 | { |
| 559 | throwConfigError(paramName, "std::size_t", |
| 560 | line, lineNum); |
| 561 | } |
| 562 | } |
| 563 | else if (paramName == "rightEyeOpenNumTaps") |
| 564 | { |
| 565 | if (!(ss >> m_cfg.rightEyeOpenNumTaps)) |
| 566 | { |
| 567 | throwConfigError(paramName, "std::size_t", |
| 568 | line, lineNum); |
| 569 | } |
| 570 | } |
| 571 | else if (paramName == "cvWaitKeyMs") |
| 572 | { |
| 573 | if (!(ss >> m_cfg.cvWaitKeyMs)) |
| 574 | { |
| 575 | throwConfigError(paramName, "int", |
| 576 | line, lineNum); |
| 577 | } |
| 578 | } |
| 579 | else if (paramName == "eyeClosedThreshold") |
| 580 | { |
| 581 | if (!(ss >> m_cfg.eyeClosedThreshold)) |
| 582 | { |
| 583 | throwConfigError(paramName, "double", |
| 584 | line, lineNum); |
| 585 | } |
| 586 | } |
| 587 | else if (paramName == "eyeOpenThreshold") |
| 588 | { |
| 589 | if (!(ss >> m_cfg.eyeOpenThreshold)) |
| 590 | { |
| 591 | throwConfigError(paramName, "double", |
| 592 | line, lineNum); |
| 593 | } |
| 594 | } |
| 595 | else if (paramName == "mouthNormalThreshold") |
| 596 | { |
| 597 | if (!(ss >> m_cfg.mouthNormalThreshold)) |
| 598 | { |
| 599 | throwConfigError(paramName, "double", |
| 600 | line, lineNum); |
| 601 | } |
| 602 | } |
| 603 | else if (paramName == "mouthSmileThreshold") |
| 604 | { |
| 605 | if (!(ss >> m_cfg.mouthSmileThreshold)) |
| 606 | { |
| 607 | throwConfigError(paramName, "double", |
| 608 | line, lineNum); |
| 609 | } |
| 610 | } |
| 611 | else if (paramName == "mouthClosedThreshold") |
| 612 | { |
| 613 | if (!(ss >> m_cfg.mouthClosedThreshold)) |
| 614 | { |
| 615 | throwConfigError(paramName, "double", |
| 616 | line, lineNum); |
| 617 | } |
| 618 | } |
| 619 | else if (paramName == "mouthOpenThreshold") |
| 620 | { |
| 621 | if (!(ss >> m_cfg.mouthOpenThreshold)) |
| 622 | { |
| 623 | throwConfigError(paramName, "double", |
| 624 | line, lineNum); |
| 625 | } |
| 626 | } |
| 627 | else if (paramName == "mouthOpenLaughCorrection") |
| 628 | { |
| 629 | if (!(ss >> m_cfg.mouthOpenLaughCorrection)) |
| 630 | { |
| 631 | throwConfigError(paramName, "double", |
| 632 | line, lineNum); |
| 633 | } |
| 634 | } |
| 635 | else if (paramName == "faceYAngleXRotCorrection") |
| 636 | { |
| 637 | if (!(ss >> m_cfg.faceYAngleXRotCorrection)) |
| 638 | { |
| 639 | throwConfigError(paramName, "double", |
| 640 | line, lineNum); |
| 641 | } |
| 642 | } |
| 643 | else if (paramName == "faceYAngleSmileCorrection") |
| 644 | { |
| 645 | if (!(ss >> m_cfg.faceYAngleSmileCorrection)) |
| 646 | { |
| 647 | throwConfigError(paramName, "double", |
| 648 | line, lineNum); |
| 649 | } |
| 650 | } |
| 651 | else if (paramName == "faceYAngleZeroValue") |
| 652 | { |
| 653 | if (!(ss >> m_cfg.faceYAngleZeroValue)) |
| 654 | { |
| 655 | throwConfigError(paramName, "double", |
| 656 | line, lineNum); |
| 657 | } |
| 658 | } |
| 659 | else if (paramName == "faceYAngleUpThreshold") |
| 660 | { |
| 661 | if (!(ss >> m_cfg.faceYAngleUpThreshold)) |
| 662 | { |
| 663 | throwConfigError(paramName, "double", |
| 664 | line, lineNum); |
| 665 | } |
| 666 | } |
| 667 | else if (paramName == "faceYAngleDownThreshold") |
| 668 | { |
| 669 | if (!(ss >> m_cfg.faceYAngleDownThreshold)) |
| 670 | { |
| 671 | throwConfigError(paramName, "double", |
| 672 | line, lineNum); |
| 673 | } |
| 674 | } |
| 675 | else if (paramName == "autoBlink") |
| 676 | { |
| 677 | if (!(ss >> m_cfg.autoBlink)) |
| 678 | { |
| 679 | throwConfigError(paramName, "bool", |
| 680 | line, lineNum); |
| 681 | } |
| 682 | } |
| 683 | else if (paramName == "autoBreath") |
| 684 | { |
| 685 | if (!(ss >> m_cfg.autoBreath)) |
| 686 | { |
| 687 | throwConfigError(paramName, "bool", |
| 688 | line, lineNum); |
| 689 | } |
| 690 | } |
| 691 | else if (paramName == "randomMotion") |
| 692 | { |
| 693 | if (!(ss >> m_cfg.randomMotion)) |
| 694 | { |
| 695 | throwConfigError(paramName, "bool", |
| 696 | line, lineNum); |
| 697 | } |
| 698 | } |
| 699 | else |
| 700 | { |
| 701 | std::ostringstream oss; |
| 702 | oss << "Unrecognized parameter name at line " << lineNum |
| 703 | << ": " << paramName; |
| 704 | throw std::runtime_error(oss.str()); |
| 705 | } |
| 706 | } |
| 707 | } |
| 708 | } |
| 709 | } |
| 710 | |
| 711 | void FacialLandmarkDetector::populateDefaultConfig(void) |
| 712 | { |
| 713 | // These are values that I've personally tested to work OK for my face. |
| 714 | // Your milage may vary - hence the config file. |
| 715 | |
| 716 | m_cfg.cvVideoCaptureId = 0; |
| 717 | m_cfg.predictorPath = "shape_predictor_68_face_landmarks.dat"; |
| 718 | m_cfg.faceYAngleCorrection = 10; |
| 719 | m_cfg.eyeSmileEyeOpenThreshold = 0.6; |
| 720 | m_cfg.eyeSmileMouthFormThreshold = 0.75; |
| 721 | m_cfg.eyeSmileMouthOpenThreshold = 0.5; |
| 722 | m_cfg.showWebcamVideo = true; |
| 723 | m_cfg.renderLandmarksOnVideo = true; |
| 724 | m_cfg.lateralInversion = true; |
| 725 | m_cfg.cvWaitKeyMs = 5; |
| 726 | m_cfg.faceXAngleNumTaps = 11; |
| 727 | m_cfg.faceYAngleNumTaps = 11; |
| 728 | m_cfg.faceZAngleNumTaps = 11; |
| 729 | m_cfg.mouthFormNumTaps = 3; |
| 730 | m_cfg.mouthOpenNumTaps = 3; |
| 731 | m_cfg.leftEyeOpenNumTaps = 3; |
| 732 | m_cfg.rightEyeOpenNumTaps = 3; |
| 733 | m_cfg.eyeClosedThreshold = 0.2; |
| 734 | m_cfg.eyeOpenThreshold = 0.25; |
| 735 | m_cfg.mouthNormalThreshold = 0.75; |
| 736 | m_cfg.mouthSmileThreshold = 1.0; |
| 737 | m_cfg.mouthClosedThreshold = 0.1; |
| 738 | m_cfg.mouthOpenThreshold = 0.4; |
| 739 | m_cfg.mouthOpenLaughCorrection = 0.2; |
| 740 | m_cfg.faceYAngleXRotCorrection = 0.15; |
| 741 | m_cfg.faceYAngleSmileCorrection = 0.075; |
| 742 | m_cfg.faceYAngleZeroValue = 1.8; |
| 743 | m_cfg.faceYAngleDownThreshold = 2.3; |
| 744 | m_cfg.faceYAngleUpThreshold = 1.3; |
| 745 | m_cfg.autoBlink = false; |
| 746 | m_cfg.autoBreath = false; |
| 747 | m_cfg.randomMotion = false; |
| 748 | } |
| 749 | |
| 750 | void FacialLandmarkDetector::throwConfigError(std::string paramName, |
| 751 | std::string expectedType, |
| 752 | std::string line, |
| 753 | unsigned int lineNum) |
| 754 | { |
| 755 | std::ostringstream ss; |
| 756 | ss << "Error parsing config file for parameter " << paramName |
| 757 | << "\nAt line " << lineNum << ": " << line |
| 758 | << "\nExpecting value of type " << expectedType; |
| 759 | |
| 760 | throw std::runtime_error(ss.str()); |
| 761 | } |
| 762 | |