def removeHotPixels(inDict, **kwargs):
# boilerplate to get down to dvs container
if isinstance(inDict, list):
return [removeHotPixels(inDictSingle, **kwargs)
for inDictSingle in inDict]
if not isinstance(inDict, dict):
return inDict
if 'ts' not in inDict:
outDict = {}
for key in inDict.keys():
outDict[key] = removeHotPixels(inDict[key], **kwargs)
return outDict
# From this point onwards, it's a data-type container
if 'pol' not in inDict:
return
# From this point onwards, it's a dvs container
events = inDict
eventImage = getEventImage(events, contrast=np.inf, polarised=False)
contrast1d = eventImage.flatten()
mean = np.mean(contrast1d)
std = np.std(contrast1d)
threshold = mean + kwargs.get('threshold', 3) * std
(y, x) = np.where(eventImage > threshold)
dimY = kwargs.get('dimY', events.get('dimY', events['y'].max() + 1))
#dimX = kwargs.get('dimX', events.get('dimX', events['x'].max()))
addrsToRemove = x * dimY + y
eventAddrs = events['x'] * dimY + events['y']
toKeep = np.logical_not(np.isin(eventAddrs, addrsToRemove))
return selectByBool(events, toKeep)
def refractoryPeriod(inDict, refractoryPeriod=0.001, **kwargs):
# boilerplate to get down to dvs container
if isinstance(inDict, list):
return [refractoryPeriod(inDictSingle, **kwargs)
for inDictSingle in inDict]
if not isinstance(inDict, dict):
return inDict
if 'ts' not in inDict:
outDict = {}
for key in inDict.keys():
outDict[key] = refractoryPeriod(inDict[key], **kwargs)
return outDict
# From this point onwards, it's a data-type container
if 'pol' not in inDict:
return
# From this point onwards, it's a dvs container
events = inDict
ts = events['ts']
x = events['y']
y = events['x']
numEvents = len(ts)
maxX = x.max()
maxY = y.max()
prevTs = np.zeros((maxY+1, maxX+1))
toKeep = np.ones((numEvents), dtype=np.bool)
for idx in trange(numEvents, leave=True, position=0):
if ts[idx] >= prevTs[y[idx], x[idx]] + refractoryPeriod:
prevTs[y[idx], x[idx]] = ts[idx]
else:
toKeep[idx] = False
outDict = selectByBool(inDict, toKeep)
return outDict
x = meshgrid[0].flatten()
y = meshgrid[1].flatten()
z = np.zeros_like(y)
# generate calibration pattern coords
coords = np.concatenate((x[:, np.newaxis], y[:, np.newaxis], z[:, np.newaxis]),
axis=1).astype(np.float32)
usableImagePoints = []
usableBool = np.zeros((len(frames['frames'])), dtype=np.bool)
for idx, frame in enumerate(tqdm(frames['frames'], file=sys.stdout)):
corners = cv2.findChessboardCorners(frames['frames'][idx],
patternSize=checkerboardDims)
if corners[0] == True:
usableBool[idx] = True
usableImagePoints.append(corners[1][:, 0, :])
frames = selectByBool(frames, usableBool)
frames['imagePoints'] = usableImagePoints
#%% Run the calibration
#frames = framesBeforeCalibration
objectPoints = [coords for idx in range(len(frames['frames']))]
# camera matrix is an initial guess
cameraMatrix = np.array([[200, 0, 152], [0, 200, 120], [0, 0, 1]],
dtype=np.float64)
distCoeffs = np.array([0, 0, 0, 0, 0], dtype=np.float64)
retval, cameraMatrix, distCoeffs, rVecs, tVecs = cv2.calibrateCamera(
objectPoints=objectPoints,
imagePoints=frames['imagePoints'],
imageSize=(240, 304),
x = meshgrid[0].flatten()
y = meshgrid[1].flatten()
z = np.zeros_like(y)
# generate calibration pattern coords
coords = np.concatenate((x[:, np.newaxis], y[:, np.newaxis], z[:, np.newaxis]),
axis=1).astype(np.float32)
usableImagePoints = []
usableBool = np.zeros((len(frames['frames'])), dtype=np.bool)
for idx, frame in enumerate(tqdm(frames['frames'])):
corners = cv2.findChessboardCorners(frames['frames'][idx],
patternSize=(7, 4))
if corners[0] == True:
usableBool[idx] = True
usableImagePoints.append(corners[1][:, 0, :])
frames = selectByBool(frames, usableBool)
frames['imagePoints'] = usableImagePoints
#%% Select poses which match the frames
poses = getSamplesAtTimes(poses, frames['ts'])
#%% Filter poses too far from frames
import math
maxTimeDifference = 0.01
numSamples = poses['ts'].shape[0]
keepBool = np.ones((numSamples), dtype=np.bool)
for idx in range(numSamples):
from bimvee.split import selectByBool
containerVicon = importAe(filePathOrName=os.path.join(filePathOrName, 'Vicon'))
# Filter out just the desired pose samples
from bimvee.importIitVicon import separateMarkersFromSegments
containerVicon['data']['vicon'] = separateMarkersFromSegments(
containerVicon['data']['vicon']['pose6q'])
del containerVicon['data']['vicon']['point3']
# Remove null pose samples (where tracking failed)
nullPoses = containerVicon['data']['vicon']['pose6q']['point'] == np.zeros(
(1, 3))
nullPoses = np.any(nullPoses, axis=1)
containerVicon['data']['vicon']['pose6q'] = selectByBool(
containerVicon['data']['vicon']['pose6q'], ~nullPoses)
#%%
from bimvee.info import info
info(containerVicon)
#%% Time alignment
'''
Now for some irreducible complexity
events and vicon were simultaneuous yarp imports. They need to be time-aligned
using tsOffsetFromInfo.
Reconstructed frames takes the same timestamp as the dvs events - they need
to be aligned according to those. Here we just use the coincidence that dvs
recording started before vicon recording, so dvs events happen to be aligned