def __runOptimization(self, problem, deltaJ, deltaX, maxIt): """Run the given optimization problem problem""" print("run new optimisation with initial values:") self.__printResults() # verbose and choldmod solving with schur complement trick options = aopt.Optimizer2Options() options.verbose = True options.nThreads = max(1, multiprocessing.cpu_count() - 1) options.doSchurComplement = True options.linearSolver = aopt.BlockCholeskyLinearSystemSolver( ) #does not have multi-threading support # stopping criteria options.maxIterations = maxIt options.convergenceDeltaJ = deltaJ options.convergenceDeltaX = deltaX # use the dogleg trustregion policy options.trustRegionPolicy = aopt.DogLegTrustRegionPolicy() # create the optimizer optimizer = aopt.Optimizer2(options) optimizer.setProblem(problem) # go for it: return optimizer.optimize()
def optimize(self, options=None, maxIterations=30, recoverCov=False): if options is None: options = aopt.Optimizer2Options() options.verbose = True options.doLevenbergMarquardt = True options.levenbergMarquardtLambdaInit = 100.0 options.nThreads = max(1,multiprocessing.cpu_count()-1) options.convergenceDeltaX = 1e-4 options.convergenceDeltaJ = 1 options.maxIterations = maxIterations options.trustRegionPolicy = aopt.LevenbergMarquardtTrustRegionPolicy(options.levenbergMarquardtLambdaInit) #run the optimization self.optimizer = aopt.Optimizer2(options) self.optimizer.setProblem(self.problem) optimizationFailed=False try: retval = self.optimizer.optimize() if retval.linearSolverFailure: optimizationFailed = True except: optimizationFailed = True if optimizationFailed: sm.logError("Optimization failed!") raise RuntimeError("Optimization failed!") #free some memory del self.optimizer gc.collect() if recoverCov: self.recoverCovariance()
def optimize(problem, options=None, maxIterations=30): if options is None: options = aopt.Optimizer2Options() options.verbose = True options.doLevenbergMarquardt = True options.levenbergMarquardtLambdaInit = 10.0 options.nThreads = max(1, multiprocessing.cpu_count() - 1) options.convergenceDeltaX = 1e-5 options.convergenceJDescentRatioThreshold = 1e-6 options.maxIterations = maxIterations options.trustRegionPolicy = aopt.LevenbergMarquardtTrustRegionPolicy( options.levenbergMarquardtLambdaInit) options.linearSolver = aopt.BlockCholeskyLinearSystemSolver() # run the optimization optimizer = aopt.Optimizer2(options) optimizer.setProblem(problem) optimizationFailed = False try: retval = optimizer.optimize() if retval.linearSolverFailure: optimizationFailed = True except: optimizationFailed = True if optimizationFailed: sm.logError("Optimization failed!") raise RuntimeError("Optimization failed!") # free some memory del optimizer gc.collect()
def maximization_step(self, t_source, target, estep_res, w=0.0, objective_type='pt2pt'): m, ndim = t_source.shape n = target.shape[0] assert ndim == 3, "ndim must be 3." m0, m1, nx = estep_res c = w / (1.0 - w) * n / m m0[m0==0] = np.finfo(np.float32).eps m1m0 = np.divide(m1.T, m0).T m0m0 = m0 / (m0 + c) drxdx = m0m0 errs = [] self.problem.clearAllErrorTerms() if objective_type == 'pt2pt': for i in xrange(m): T_l_l = self.T_b_l_Dv.toExpression().inverse() * \ aopt.TransformationExpression(self._sensor_tfs[i]) * \ self.T_b_l_Dv.toExpression() predicted = T_l_l.toRotationExpression() * t_source[i] + T_l_l.toEuclideanExpression() err = ket.EuclideanError(m1m0[i], drxdx[i]*np.eye(3, dtype=np.float64), predicted) errs.append(err) self.problem.addErrorTerm(err) else: raise ValueError('Unknown objective_type: %s.' % objective_type) # define the optimization options = aopt.Optimizer2Options() options.verbose = False options.linearSolver = aopt.BlockCholeskyLinearSystemSolver() options.nThreads = 2 options.convergenceDeltaX = 1e-4 options.convergenceJDescentRatioThreshold = 1e-6 options.maxIterations = 50 # run the optimization optimizer = aopt.Optimizer2(options) optimizer.setProblem(self.problem) # get the prior try: optimizer.optimize() except: sm.logFatal("Failed to obtain orientation prior!") sys.exit(-1) q = np.array([np.linalg.norm(e.error()) for e in errs]).sum() return MstepResult(self.T_b_l_Dv.toExpression().toTransformationMatrix(), q)
def findOrientationPriorCameraToImu(self, imu): print print "Estimating imu-camera rotation prior" # build the problem problem = aopt.OptimizationProblem() # Add the rotation as design variable q_i_c_Dv = aopt.RotationQuaternionDv(self.T_extrinsic.q()) q_i_c_Dv.setActive(True) problem.addDesignVariable(q_i_c_Dv) # Add the gyro bias as design variable gyroBiasDv = aopt.EuclideanPointDv(np.zeros(3)) gyroBiasDv.setActive(True) problem.addDesignVariable(gyroBiasDv) #initialize a pose spline using the camera poses poseSpline = self.initPoseSplineFromCamera(timeOffsetPadding=0.0) for im in imu.imuData: tk = im.stamp.toSec() if tk > poseSpline.t_min() and tk < poseSpline.t_max(): #DV expressions R_i_c = q_i_c_Dv.toExpression() bias = gyroBiasDv.toExpression() #get the vision predicted omega and measured omega (IMU) omega_predicted = R_i_c * aopt.EuclideanExpression( np.matrix( poseSpline.angularVelocityBodyFrame(tk)).transpose()) omega_measured = im.omega #error term gerr = ket.GyroscopeError(omega_measured, im.omegaInvR, omega_predicted, bias) problem.addErrorTerm(gerr) #define the optimization options = aopt.Optimizer2Options() options.verbose = False options.linearSolver = aopt.BlockCholeskyLinearSystemSolver() options.nThreads = 2 options.convergenceDeltaX = 1e-4 options.convergenceDeltaJ = 1 options.maxIterations = 50 #run the optimization optimizer = aopt.Optimizer2(options) optimizer.setProblem(problem) #get the prior try: optimizer.optimize() except: sm.logFatal("Failed to obtain orientation prior!") sys.exit(-1) #overwrite the external rotation prior (keep the external translation prior) R_i_c = q_i_c_Dv.toRotationMatrix().transpose() self.T_extrinsic = sm.Transformation( sm.rt2Transform(R_i_c, self.T_extrinsic.t())) #set the gyro bias prior (if we have more than 1 cameras use recursive average) b_gyro = bias.toEuclidean() imu.GyroBiasPriorCount += 1 imu.GyroBiasPrior = ( imu.GyroBiasPriorCount - 1.0 ) / imu.GyroBiasPriorCount * imu.GyroBiasPrior + 1.0 / imu.GyroBiasPriorCount * b_gyro #print result print " Orientation prior camera-imu found as: (T_i_c)" print R_i_c print " Gyro bias prior found as: (b_gyro)" print b_gyro
def solveFullBatch(cameras, baseline_guesses, graph): ############################################ ## solve the bundle adjustment ############################################ problem = aopt.OptimizationProblem() #add camera dvs for cam in cameras: cam.setDvActiveStatus(True, True, False) problem.addDesignVariable(cam.dv.distortionDesignVariable()) problem.addDesignVariable(cam.dv.projectionDesignVariable()) problem.addDesignVariable(cam.dv.shutterDesignVariable()) baseline_dvs = list() for baseline_idx in range(0, len(cameras) - 1): baseline_dv = aopt.TransformationDv(baseline_guesses[baseline_idx]) for i in range(0, baseline_dv.numDesignVariables()): problem.addDesignVariable(baseline_dv.getDesignVariable(i)) baseline_dvs.append(baseline_dv) #corner uncertainty cornerUncertainty = 1.0 R = np.eye(2) * cornerUncertainty * cornerUncertainty invR = np.linalg.inv(R) #get the target target = cameras[0].ctarget.detector.target() #Add calibration target reprojection error terms for all camera in chain target_pose_dvs = list() #shuffle the views reprojectionErrors = [] timestamps = graph.obs_db.getAllViewTimestamps() for view_id, timestamp in enumerate(timestamps): #get all observations for all cams at this time obs_tuple = graph.obs_db.getAllObsAtTimestamp(timestamp) #create a target pose dv for all target views (= T_cam0_w) T0 = graph.getTargetPoseGuess(timestamp, cameras, baseline_guesses) target_pose_dv = addPoseDesignVariable(problem, T0) target_pose_dvs.append(target_pose_dv) for cidx, obs in obs_tuple: cam = cameras[cidx] #calibration target coords to camera X coords T_cam0_calib = target_pose_dv.toExpression().inverse() #build pose chain (target->cam0->baselines->camN) T_camN_calib = T_cam0_calib for idx in range(0, cidx): T_camN_calib = baseline_dvs[idx].toExpression() * T_camN_calib ## add error terms for i in range(0, target.size()): p_target = aopt.HomogeneousExpression( sm.toHomogeneous(target.point(i))) valid, y = obs.imagePoint(i) if valid: rerr = cameras[cidx].model.reprojectionError( y, invR, T_camN_calib * p_target, cameras[cidx].dv) problem.addErrorTerm(rerr) reprojectionErrors.append(rerr) sm.logDebug("solveFullBatch: added {0} camera error terms".format( len(reprojectionErrors))) ############################################ ## solve ############################################ options = aopt.Optimizer2Options() options.verbose = True if sm.getLoggingLevel( ) == sm.LoggingLevel.Debug else False options.nThreads = 4 options.convergenceDeltaX = 1e-3 options.convergenceDeltaJ = 1 options.maxIterations = 250 options.trustRegionPolicy = aopt.LevenbergMarquardtTrustRegionPolicy(10) optimizer = aopt.Optimizer2(options) optimizer.setProblem(problem) #verbose output if sm.getLoggingLevel() == sm.LoggingLevel.Debug: sm.logDebug("Before optimization:") e2 = np.array([e.evaluateError() for e in reprojectionErrors]) sm.logDebug( " Reprojection error squarred (camL): mean {0}, median {1}, std: {2}" .format(np.mean(e2), np.median(e2), np.std(e2))) #run intrinsic calibration try: retval = optimizer.optimize() if retval.linearSolverFailure: sm.logError("calibrateIntrinsics: Optimization failed!") success = not retval.linearSolverFailure except: sm.logError("calibrateIntrinsics: Optimization failed!") success = False baselines = list() for baseline_dv in baseline_dvs: baselines.append(sm.Transformation(baseline_dv.T())) return success, baselines
def calibrateIntrinsics(cam_geometry, obslist, distortionActive=True, intrinsicsActive=True): #verbose output if sm.getLoggingLevel() == sm.LoggingLevel.Debug: d = cam_geometry.geometry.projection().distortion().getParameters( ).flatten() p = cam_geometry.geometry.projection().getParameters().flatten() sm.logDebug("calibrateIntrinsics: intrinsics guess: {0}".format(p)) sm.logDebug("calibrateIntrinsics: distortion guess: {0}".format(d)) ############################################ ## solve the bundle adjustment ############################################ problem = aopt.OptimizationProblem() #add camera dvs cam_geometry.setDvActiveStatus(intrinsicsActive, distortionActive, False) problem.addDesignVariable(cam_geometry.dv.distortionDesignVariable()) problem.addDesignVariable(cam_geometry.dv.projectionDesignVariable()) problem.addDesignVariable(cam_geometry.dv.shutterDesignVariable()) #corner uncertainty cornerUncertainty = 1.0 R = np.eye(2) * cornerUncertainty * cornerUncertainty invR = np.linalg.inv(R) #get the image and target points corresponding to the frame target = cam_geometry.ctarget.detector.target() #target pose dv for all target views (=T_camL_w) reprojectionErrors = [] sm.logDebug( "calibrateIntrinsics: adding camera error terms for {0} calibration targets" .format(len(obslist))) target_pose_dvs = list() for obs in obslist: success, T_t_c = cam_geometry.geometry.estimateTransformation(obs) target_pose_dv = addPoseDesignVariable(problem, T_t_c) target_pose_dvs.append(target_pose_dv) T_cam_w = target_pose_dv.toExpression().inverse() ## add error terms for i in range(0, target.size()): p_target = aopt.HomogeneousExpression( sm.toHomogeneous(target.point(i))) valid, y = obs.imagePoint(i) if valid: rerr = cam_geometry.model.reprojectionError( y, invR, T_cam_w * p_target, cam_geometry.dv) problem.addErrorTerm(rerr) reprojectionErrors.append(rerr) sm.logDebug("calibrateIntrinsics: added {0} camera error terms".format( len(reprojectionErrors))) ############################################ ## solve ############################################ options = aopt.Optimizer2Options() options.verbose = True if sm.getLoggingLevel( ) == sm.LoggingLevel.Debug else False options.nThreads = 4 options.convergenceDeltaX = 1e-3 options.convergenceDeltaJ = 1 options.maxIterations = 200 options.trustRegionPolicy = aopt.LevenbergMarquardtTrustRegionPolicy(10) optimizer = aopt.Optimizer2(options) optimizer.setProblem(problem) #verbose output if sm.getLoggingLevel() == sm.LoggingLevel.Debug: sm.logDebug("Before optimization:") e2 = np.array([e.evaluateError() for e in reprojectionErrors]) sm.logDebug( " Reprojection error squarred (camL): mean {0}, median {1}, std: {2}" .format(np.mean(e2), np.median(e2), np.std(e2))) #run intrinsic calibration try: retval = optimizer.optimize() if retval.linearSolverFailure: sm.logError("calibrateIntrinsics: Optimization failed!") success = not retval.linearSolverFailure except: sm.logError("calibrateIntrinsics: Optimization failed!") success = False #verbose output if sm.getLoggingLevel() == sm.LoggingLevel.Debug: d = cam_geometry.geometry.projection().distortion().getParameters( ).flatten() p = cam_geometry.geometry.projection().getParameters().flatten() sm.logDebug( "calibrateIntrinsics: guess for intrinsics cam: {0}".format(p)) sm.logDebug( "calibrateIntrinsics: guess for distortion cam: {0}".format(d)) return success
def stereoCalibrate(camL_geometry, camH_geometry, obslist, distortionActive=False, baseline=None): ##################################################### ## find initial guess as median of all pnp solutions ##################################################### if baseline is None: r = [] t = [] for obsL, obsH in obslist: #if we have observations for both camss if obsL is not None and obsH is not None: success, T_L = camL_geometry.geometry.estimateTransformation( obsL) success, T_H = camH_geometry.geometry.estimateTransformation( obsH) baseline = T_H.inverse() * T_L t.append(baseline.t()) rv = sm.RotationVector() r.append(rv.rotationMatrixToParameters(baseline.C())) r_median = np.median(np.asmatrix(r), axis=0).flatten().T R_median = rv.parametersToRotationMatrix(r_median) t_median = np.median(np.asmatrix(t), axis=0).flatten().T baseline_HL = sm.Transformation(sm.rt2Transform(R_median, t_median)) else: baseline_HL = baseline #verbose output if sm.getLoggingLevel() == sm.LoggingLevel.Debug: dL = camL_geometry.geometry.projection().distortion().getParameters( ).flatten() pL = camL_geometry.geometry.projection().getParameters().flatten() dH = camH_geometry.geometry.projection().distortion().getParameters( ).flatten() pH = camH_geometry.geometry.projection().getParameters().flatten() sm.logDebug("initial guess for stereo calib: {0}".format( baseline_HL.T())) sm.logDebug("initial guess for intrinsics camL: {0}".format(pL)) sm.logDebug("initial guess for intrinsics camH: {0}".format(pH)) sm.logDebug("initial guess for distortion camL: {0}".format(dL)) sm.logDebug("initial guess for distortion camH: {0}".format(dH)) ############################################ ## solve the bundle adjustment ############################################ problem = aopt.OptimizationProblem() #baseline design variable baseline_dv = addPoseDesignVariable(problem, baseline_HL) #target pose dv for all target views (=T_camL_w) target_pose_dvs = list() for obsL, obsH in obslist: if obsL is not None: #use camL if we have an obs for this one success, T_t_cL = camL_geometry.geometry.estimateTransformation( obsL) else: success, T_t_cH = camH_geometry.geometry.estimateTransformation( obsH) T_t_cL = T_t_cH * baseline_HL #apply baseline for the second camera target_pose_dv = addPoseDesignVariable(problem, T_t_cL) target_pose_dvs.append(target_pose_dv) #add camera dvs camL_geometry.setDvActiveStatus(True, distortionActive, False) camH_geometry.setDvActiveStatus(True, distortionActive, False) problem.addDesignVariable(camL_geometry.dv.distortionDesignVariable()) problem.addDesignVariable(camL_geometry.dv.projectionDesignVariable()) problem.addDesignVariable(camL_geometry.dv.shutterDesignVariable()) problem.addDesignVariable(camH_geometry.dv.distortionDesignVariable()) problem.addDesignVariable(camH_geometry.dv.projectionDesignVariable()) problem.addDesignVariable(camH_geometry.dv.shutterDesignVariable()) ############################################ ## add error terms ############################################ #corner uncertainty # \todo pass in the detector uncertainty somehow. cornerUncertainty = 1.0 R = np.eye(2) * cornerUncertainty * cornerUncertainty invR = np.linalg.inv(R) #Add reprojection error terms for both cameras reprojectionErrors0 = [] reprojectionErrors1 = [] for cidx, cam in enumerate([camL_geometry, camH_geometry]): sm.logDebug( "stereoCalibration: adding camera error terms for {0} calibration targets" .format(len(obslist))) #get the image and target points corresponding to the frame target = cam.ctarget.detector.target() #add error terms for all observations for view_id, obstuple in enumerate(obslist): #add error terms if we have an observation for this cam obs = obstuple[cidx] if obs is not None: T_cam_w = target_pose_dvs[view_id].toExpression().inverse() #add the baseline for the second camera if cidx != 0: T_cam_w = baseline_dv.toExpression() * T_cam_w for i in range(0, target.size()): p_target = aopt.HomogeneousExpression( sm.toHomogeneous(target.point(i))) valid, y = obs.imagePoint(i) if valid: # Create an error term. rerr = cam.model.reprojectionError( y, invR, T_cam_w * p_target, cam.dv) rerr.idx = i problem.addErrorTerm(rerr) if cidx == 0: reprojectionErrors0.append(rerr) else: reprojectionErrors1.append(rerr) sm.logDebug("stereoCalibrate: added {0} camera error terms".format( len(reprojectionErrors0) + len(reprojectionErrors1))) ############################################ ## solve ############################################ options = aopt.Optimizer2Options() options.verbose = True if sm.getLoggingLevel( ) == sm.LoggingLevel.Debug else False options.nThreads = 4 options.convergenceDeltaX = 1e-3 options.convergenceDeltaJ = 1 options.maxIterations = 200 options.trustRegionPolicy = aopt.LevenbergMarquardtTrustRegionPolicy(10) optimizer = aopt.Optimizer2(options) optimizer.setProblem(problem) #verbose output if sm.getLoggingLevel() == sm.LoggingLevel.Debug: sm.logDebug("Before optimization:") e2 = np.array([e.evaluateError() for e in reprojectionErrors0]) sm.logDebug( " Reprojection error squarred (camL): mean {0}, median {1}, std: {2}" .format(np.mean(e2), np.median(e2), np.std(e2))) e2 = np.array([e.evaluateError() for e in reprojectionErrors1]) sm.logDebug( " Reprojection error squarred (camH): mean {0}, median {1}, std: {2}" .format(np.mean(e2), np.median(e2), np.std(e2))) sm.logDebug("baseline={0}".format( baseline_dv.toTransformationMatrix())) try: retval = optimizer.optimize() if retval.linearSolverFailure: sm.logError("stereoCalibrate: Optimization failed!") success = not retval.linearSolverFailure except: sm.logError("stereoCalibrate: Optimization failed!") success = False if sm.getLoggingLevel() == sm.LoggingLevel.Debug: sm.logDebug("After optimization:") e2 = np.array([e.evaluateError() for e in reprojectionErrors0]) sm.logDebug( " Reprojection error squarred (camL): mean {0}, median {1}, std: {2}" .format(np.mean(e2), np.median(e2), np.std(e2))) e2 = np.array([e.evaluateError() for e in reprojectionErrors1]) sm.logDebug( " Reprojection error squarred (camH): mean {0}, median {1}, std: {2}" .format(np.mean(e2), np.median(e2), np.std(e2))) #verbose output if sm.getLoggingLevel() == sm.LoggingLevel.Debug: dL = camL_geometry.geometry.projection().distortion().getParameters( ).flatten() pL = camL_geometry.geometry.projection().getParameters().flatten() dH = camH_geometry.geometry.projection().distortion().getParameters( ).flatten() pH = camH_geometry.geometry.projection().getParameters().flatten() sm.logDebug("guess for intrinsics camL: {0}".format(pL)) sm.logDebug("guess for intrinsics camH: {0}".format(pH)) sm.logDebug("guess for distortion camL: {0}".format(dL)) sm.logDebug("guess for distortion camH: {0}".format(dH)) if success: baseline_HL = sm.Transformation(baseline_dv.toTransformationMatrix()) return success, baseline_HL else: #return the initial guess if we fail return success, baseline_HL