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 addTargetView(self, rig_observations, T_tc_guess, force=False):
#create the problem for this batch and try to add it
batch_problem = CalibrationTargetOptimizationProblem.fromTargetViewObservations(self.cameras, self.target, self.baselines, T_tc_guess, rig_observations, useBlakeZissermanMest=self.useBlakeZissermanMest)
self.estimator_return_value = self.estimator.addBatch(batch_problem, force)
if self.estimator_return_value.numIterations >= self.optimizerOptions.maxIterations:
sm.logError("Did not converge in maxIterations... restarting...")
raise OptimizationDiverged
success = self.estimator_return_value.batchAccepted
if success:
sm.logDebug("The estimator accepted this batch")
self.views.append(batch_problem)
else:
sm.logDebug("The estimator did not accept this batch")
return success
def addObservation(self, cam_id, obs):
#create camera list (initialization)
if cam_id not in self.observations:
self.observations[cam_id] = list()
#add to archive
self.observations[cam_id].append(obs)
obs_idx = len(self.observations[cam_id])-1
#find the nearest timestamp in the table
timestamps_table = self.targetViews.keys()
timestamp_obs = obs.time().toSec()
#check if the table is still empty (initialization)
if not timestamps_table:
#force a new entry (by putting the "nearest" timestamp more than max_delta_approxsync away)
nearest_timestamp = timestamp_obs + 5*(self.max_delta_approxsync+1)
else:
nearest_timestamp = min(timestamps_table, key=lambda x: abs(x-timestamp_obs))
#if +-max approx. sync add to this time instant otherwise create a new timestamp)
if abs(nearest_timestamp-timestamp_obs) <= self.max_delta_approxsync:
#add to existing timestamp
timestamp = nearest_timestamp
else:
#add new timestamp
timestamp = timestamp_obs
self.targetViews[ timestamp ] = dict()
#fill in observation data
if cam_id not in self.targetViews[timestamp]:
#create entry if it doesnt exists
self.targetViews[timestamp][cam_id] = dict()
self.targetViews[timestamp][cam_id]['obs_id'] = obs_idx
self.targetViews[timestamp][cam_id]['observed_corners'] = set(obs.getCornersIdx())
else:
#we already have a view from this camera on this timestamp --> STH IS WRONG
sm.logError("[TargetViewTable]: Tried to add second view to a given cameraId & "
"timestamp. Maybe try to reduce the approximate syncing tolerance..")
def getAllMutualObsBetweenTwoCams(self, camA_nr, camB_nr):
#get the observation ids
try:
edge_idx = self.G.get_eid(camA_nr, camB_nr)
except:
sm.logError("getAllMutualObsBetweenTwoCams: no mutual observations between the two cams!")
return [], []
observations = self.G.es[edge_idx]["obs_ids"]
#extract the ids
obs_idx_L = [obs_ids[0] for obs_ids in observations]
obs_idx_H = [obs_ids[1] for obs_ids in observations]
#the first value of the tuple always stores the obsvervations for camera
#with the lower id
obs_idx_A = obs_idx_L if camA_nr<camB_nr else obs_idx_H
obs_idx_B = obs_idx_H if camA_nr<camB_nr else obs_idx_L
#get the obs from the storage using idx
obs_A = [self.obs_db.observations[camA_nr][idx] for idx in obs_idx_A]
obs_B = [self.obs_db.observations[camB_nr][idx] for idx in obs_idx_B]
return obs_A, obs_B
def initGeometryFromObservations(self, observations):
#obtain focal length guess
success = self.geometry.initializeIntrinsics(observations)
if not success:
sm.logError("initialization of focal length for cam with topic {0} failed ".format(self.dataset.topic))
#in case of an omni model, first optimize over intrinsics only
#(--> catch most of the distortion with the projection model)
if self.model == acvb.DistortedOmni:
success = kcc.calibrateIntrinsics(self, observations, distortionActive=False)
if not success:
sm.logError("initialization of intrinsics for cam with topic {0} failed ".format(self.dataset.topic))
#optimize for intrinsics & distortion
success = kcc.calibrateIntrinsics(self, observations)
if not success:
sm.logError("initialization of intrinsics for cam with topic {0} failed ".format(self.dataset.topic))
self.isGeometryInitialized = success
return success
def calibrate(self,
cameraGeometry,
observations,
config
):
"""
A Motion regularization term is added with low a priori knowledge to avoid
diverging parts in the spline of too many knots are selected/provided or if
no image information is available for long sequences and to regularize the
last few frames (which typically contain no image information but need to have
knots to /close/ the spline).
Kwargs:
cameraGeometry (kcc.CameraGeometry): a camera geometry object with an initialized target
observations ([]: The list of observation \see extractCornersFromDataset
config (RsCalibratorConfiguration): calibration configuration
"""
## set internal objects
self.__observations = observations
self.__cameraGeometry = cameraGeometry
self.__cameraModelFactory = cameraGeometry.model
self.__camera_dv = cameraGeometry.dv
self.__camera = cameraGeometry.geometry
self.__config = config
self.__config.validate(self.__isRollingShutter())
# obtain initial guesses for extrinsics and intrinsics
if (not self.__generateIntrinsicsInitialGuess()):
sm.logError("Could not generate initial guess.")
# obtain the extrinsic initial guess for every observation
self.__generateExtrinsicsInitialGuess()
# set the value for the motion prior term or uses the defaults
W = self.__getMotionModelPriorOrDefault()
self.__poseSpline = self.__generateInitialSpline(
self.__config.splineOrder,
self.__config.timeOffsetPadding,
self.__config.numberOfKnots,
self.__config.framerate
)
# build estimator problem
optimisation_problem = self.__buildOptimizationProblem(W)
self.__runOptimization(
optimisation_problem,
self.__config.deltaJ,
self.__config.deltaX,
self.__config.maxNumberOfIterations
)
# continue with knot replacement
if self.__config.adaptiveKnotPlacement:
knotUpdateStrategy = self.__config.knotUpdateStrategy(self.__config.framerate)
for iteration in range(self.__config.maxKnotPlacementIterations):
# generate the new knots list
[knots, requiresUpdate] = knotUpdateStrategy.generateKnotList(
self.__reprojection_errors,
self.__poseSpline_dv.spline()
)
# if no new knotlist was generated, we are done.
if (not requiresUpdate):
break;
# otherwise update the spline dv and rebuild the problem
self.__poseSpline = knotUpdateStrategy.getUpdatedSpline(self.__poseSpline_dv.spline(), knots, self.__config.splineOrder)
optimisation_problem = self.__buildOptimizationProblem(W)
self.__runOptimization(
optimisation_problem,
self.__config.deltaJ,
self.__config.deltaX,
self.__config.maxNumberOfIterations
)
self.__printResults()
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 intiial guess if we fail
return success, baseline_HL
def calibrate(self, cameraGeometry, observations, config):
"""
A Motion regularization term is added with low a priori knowledge to avoid
diverging parts in the spline of too many knots are selected/provided or if
no image information is available for long sequences and to regularize the
last few frames (which typically contain no image information but need to have
knots to /close/ the spline).
Kwargs:
cameraGeometry (kcc.CameraGeometry): a camera geometry object with an initialized target
observations ([]: The list of observation \see extractCornersFromDataset
config (RsCalibratorConfiguration): calibration configuration
"""
## set internal objects
self.__observations = observations
self.__cameraGeometry = cameraGeometry
self.__cameraModelFactory = cameraGeometry.model
self.__camera_dv = cameraGeometry.dv
self.__camera = cameraGeometry.geometry
self.__config = config
self.__config.validate(self.__isRollingShutter())
# obtain initial guesses for extrinsics and intrinsics
if (not self.__generateIntrinsicsInitialGuess()):
sm.logError("Could not generate initial guess.")
# obtain the extrinsic initial guess for every observation
self.__generateExtrinsicsInitialGuess()
# set the value for the motion prior term or uses the defaults
W = self.__getMotionModelPriorOrDefault()
self.__poseSpline = self.__generateInitialSpline(
self.__config.splineOrder, self.__config.timeOffsetPadding,
self.__config.numberOfKnots, self.__config.framerate)
# build estimator problem
optimisation_problem = self.__buildOptimizationProblem(W)
self.__runOptimization(optimisation_problem, self.__config.deltaJ,
self.__config.deltaX,
self.__config.maxNumberOfIterations)
# continue with knot replacement
if self.__config.adaptiveKnotPlacement:
knotUpdateStrategy = self.__config.knotUpdateStrategy(
self.__config.framerate)
for iteration in range(self.__config.maxKnotPlacementIterations):
# generate the new knots list
[knots, requiresUpdate] = knotUpdateStrategy.generateKnotList(
self.__reprojection_errors, self.__poseSpline_dv.spline())
# if no new knotlist was generated, we are done.
if (not requiresUpdate):
break
# otherwise update the spline dv and rebuild the problem
self.__poseSpline = knotUpdateStrategy.getUpdatedSpline(
self.__poseSpline_dv.spline(), knots,
self.__config.splineOrder)
optimisation_problem = self.__buildOptimizationProblem(W)
self.__runOptimization(optimisation_problem,
self.__config.deltaJ,
self.__config.deltaX,
self.__config.maxNumberOfIterations)
self.__printResults()
def getInitialGuesses(self, cameras):
if not self.G:
raise RuntimeError("Graph is uninitialized!")
#################################################################
## STEP 0: check if all cameras in the chain are connected
## through common target point observations
## (=all vertices connected?)
#################################################################
if not self.isGraphConnected():
sm.logError("The cameras are not connected through mutual target observations! "
"Please provide another dataset...")
self.plotGraph()
sys.exit(0)
#################################################################
## STEP 1: get baseline initial guesses by calibrating good
## camera pairs using a stereo calibration
##
#################################################################
#first we need to find the best camera pairs to obtain the initial guesses
#--> use the pairs that share the most common observed target corners
#The graph is built with weighted edges that represent the number of common
#target corners, so we can use dijkstras algorithm to get the best pair
#configuration for the initial pair calibrations
weights = [1.0/commonPoints for commonPoints in self.G.es["weight"]]
#choose the cam with the least edges as base_cam
outdegrees = self.G.vs.outdegree()
base_cam_id = outdegrees.index(min(outdegrees))
#solve for shortest path (=optimal transformation chaining)
edges_on_path = self.G.get_shortest_paths(0, weights=weights, output="epath")
self.optimal_baseline_edges = set([item for sublist in edges_on_path for item in sublist])
#################################################################
## STEP 2: solve stereo calibration problem for the baselines
## (baselines are always from lower_id to higher_id cams!)
#################################################################
#calibrate all cameras in pairs
for baseline_edge_id in self.optimal_baseline_edges:
#get the cam_nrs from the graph edge (calibrate from low to high id)
vertices = self.G.es[baseline_edge_id].tuple
if vertices[0]<vertices[1]:
camL_nr = vertices[0]
camH_nr = vertices[1]
else:
camL_nr = vertices[1]
camH_nr = vertices[0]
print "\t initializing camera pair ({0},{1})... ".format(camL_nr, camH_nr)
#run the pair extrinsic calibration
obs_list = self.obs_db.getAllObsTwoCams(camL_nr, camH_nr)
success, baseline_HL = kcc.stereoCalibrate(cameras[camL_nr],
cameras[camH_nr],
obs_list,
distortionActive=False)
if success:
sm.logDebug("baseline_{0}_{1}={2}".format(camL_nr, camH_nr, baseline_HL.T()))
else:
sm.logError("initialization of camera pair ({0},{1}) failed ".format(camL_nr, camH_nr))
sm.logError("estimated baseline_{0}_{1}={2}".format(camL_nr, camH_nr, baseline_HL.T()))
#store the baseline in the graph
self.G.es[ self.G.get_eid(camL_nr, camH_nr) ]["baseline_HL"] = baseline_HL
#################################################################
## STEP 3: transform from the "optimal" baseline chain to camera chain ordering
## (=> baseline_0 = T_c1_c0 |
#################################################################
#construct the optimal path graph
G_optimal_baselines = self.G.copy()
eid_not_optimal_path = set(range(0,len(G_optimal_baselines.es)))
for eid in self.optimal_baseline_edges:
eid_not_optimal_path.remove(eid)
G_optimal_baselines.delete_edges( eid_not_optimal_path )
#now we convert the arbitary baseline graph to baselines starting from
# cam0 and traverse the chain (cam0->cam1->cam2->camN)
baselines = []
for baseline_id in range(0, self.numCams-1):
#find the shortest path on the graph
path = G_optimal_baselines.get_shortest_paths(baseline_id, baseline_id+1)[0]
#get the baseline from cam with id baseline_id to baseline_id+1
baseline_HL = sm.Transformation()
for path_idx in range(0, len(path)-1):
source_vert = path[path_idx]
target_vert = path[path_idx+1]
T_edge = self.G.es[ self.G.get_eid(source_vert, target_vert) ]["baseline_HL"]
#correct the direction (baselines always from low to high cam id!)
T_edge = T_edge if source_vert<target_vert else T_edge.inverse()
#chain up
baseline_HL = T_edge * baseline_HL
#store in graph
baselines.append(baseline_HL)
#################################################################
## STEP 4: refine guess in full batch
#################################################################
success, baselines = kcc.solveFullBatch(cameras, baselines, self)
if not success:
sm.logWarn("Full batch refinement failed!")
return baselines
def getInitialGuesses(self, cameras):
if not self.G:
raise RuntimeError("Graph is uninitialized!")
#################################################################
## STEP 0: check if all cameras in the chain are connected
## through common target point observations
## (=all vertices connected?)
#################################################################
if not self.isGraphConnected():
sm.logError("The cameras are not connected through mutual target observations! "
"Please provide another dataset...")
self.plotGraph()
sys.exit(0)
#################################################################
## STEP 1: get baseline initial guesses by calibrating good
## camera pairs using a stereo calibration
##
#################################################################
#first we need to find the best camera pairs to obtain the initial guesses
#--> use the pairs that share the most common observed target corners
#The graph is built with weighted edges that represent the number of common
#target corners, so we can use dijkstras algorithm to get the best pair
#configuration for the initial pair calibrations
weights = [1.0/commonPoints for commonPoints in self.G.es["weight"]]
#choose the cam with the least edges as base_cam
outdegrees = self.G.vs.outdegree()
base_cam_id = outdegrees.index(min(outdegrees))
#solve for shortest path (=optimal transformation chaining)
edges_on_path = self.G.get_shortest_paths(0, weights=weights, output="epath")
self.optimal_baseline_edges = set([item for sublist in edges_on_path for item in sublist])
#################################################################
## STEP 2: solve stereo calibration problem for the baselines
## (baselines are always from lower_id to higher_id cams!)
#################################################################
#calibrate all cameras in pairs
for baseline_edge_id in self.optimal_baseline_edges:
#get the cam_nrs from the graph edge (calibrate from low to high id)
vertices = self.G.es[baseline_edge_id].tuple
if vertices[0]<vertices[1]:
camL_nr = vertices[0]
camH_nr = vertices[1]
else:
camL_nr = vertices[1]
camH_nr = vertices[0]
print "\t initializing camera pair ({0},{1})... ".format(camL_nr, camH_nr)
#run the pair extrinsic calibration
obs_list = self.obs_db.getAllObsTwoCams(camL_nr, camH_nr)
success, baseline_HL = kcc.stereoCalibrate(cameras[camL_nr],
cameras[camH_nr],
obs_list,
distortionActive=False)
if success:
sm.logDebug("baseline_{0}_{1}={2}".format(camL_nr, camH_nr, baseline_HL.T()))
else:
sm.logError("initialization of camera pair ({0},{1}) failed ".format(camL_nr, camH_nr))
sm.logError("estimated baseline_{0}_{1}={2}".format(camL_nr, camH_nr, baseline_HL.T()))
#store the baseline in the graph
self.G.es[ self.G.get_eid(camL_nr, camH_nr) ]["baseline_HL"] = baseline_HL
#################################################################
## STEP 3: transform from the "optimal" baseline chain to camera chain ordering
## (=> baseline_0 = T_c1_c0 |
#################################################################
#construct the optimal path graph
G_optimal_baselines = self.G.copy()
eid_not_optimal_path = set(range(0,len(G_optimal_baselines.es)))
for eid in self.optimal_baseline_edges:
eid_not_optimal_path.remove(eid)
G_optimal_baselines.delete_edges( eid_not_optimal_path )
#now we convert the arbitary baseline graph to baselines starting from
# cam0 and traverse the chain (cam0->cam1->cam2->camN)
baselines = []
for baseline_id in range(0, self.numCams-1):
#find the shortest path on the graph
path = G_optimal_baselines.get_shortest_paths(baseline_id, baseline_id+1)[0]
#get the baseline from cam with id baseline_id to baseline_id+1
baseline_HL = sm.Transformation()
for path_idx in range(0, len(path)-1):
source_vert = path[path_idx]
target_vert = path[path_idx+1]
T_edge = self.G.es[ self.G.get_eid(source_vert, target_vert) ]["baseline_HL"]
#correct the direction (baselines always from low to high cam id!)
T_edge = T_edge if source_vert<target_vert else T_edge.inverse()
#chain up
baseline_HL = T_edge * baseline_HL
#store in graph
baselines.append(baseline_HL)
#################################################################
## STEP 4: refine guess in full batch
#################################################################
success, baselines = kcc.solveFullBatch(cameras, baselines, self)
if not success:
sm.logWarn("Full batch refinement failed!")
return baselines