class RGBSensorNetwork(Sensor):
"""A network of cameras based on the 'real' camera network.
Args:
data_dir (str): Path to reconstruction data.
measurements (list of arrays): The measurements. These are currently used only
for masking nan pixels.
ignore_angles (optional, bool): Ignore the THETA/PHI pixel angles.
ignore_sunblock (optional, bool): Ignore the (gaussian) sun mask.
subsample (optional, int): Subsampling period.
eleveation_weight (optional, float): Weight down the horizon.
mask_measurement (optional, float): Mask measurements higher than this value.
erode_mask (optional, int): Erode the mask.
weights_color (tuple): Weight colors differently.
"""
def __init__(
self,
data_dir,
indices=None,
measurements=None,
ignore_angles=False,
ignore_sunblock=False,
subsample=-1,
elevation_weight=-1,
mask_measurement=-1,
erode_mask=-1,
weights_color=(1, 1, 1),
*args,
**kwds):
super(RGBSensorNetwork, self).__init__(*args, **kwds)
with open(os.path.join(data_dir, 'export_data.pkl'), 'rb') as f:
datas = cPickle.load(f)
if subsample < 1:
#
# For some strange reason, if I don't subsample (even by 1),
# then all shadom directions are transposed.
#
subsample = 1
logging.warn("Subsampling by 1 to avoid strange bug in directions.")
camera_ids = sorted(datas.keys())
if indices is None:
indices = range(len(camera_ids))
for id_index in indices:
cam_id = camera_ids[id_index]
for color_channel in range(3):
data = datas[cam_id]
extra_data = data["extra_data"]
if ignore_angles:
Y_shdom, X_shdom = np.meshgrid(
np.linspace(-1, 1, 301),
np.linspace(-1, 1, 301)
)
PHI = math.pi + np.arctan2(Y_shdom, X_shdom)
PSI = -math.pi + math.pi/2 * np.sqrt(X_shdom**2 + Y_shdom**2)
else:
PHI = data['PHI']
PSI = data['PSI']
#
# Prepare the MASK (erode it if requested).
#
MASK = data["MASK"].astype(np.uint8)
MASK[MASK<1] = 0
MASK[MASK>1] = 1
if erode_mask > 0:
kernel = np.ones((3, 3), np.uint8)
MASK = cv2.erode(MASK, kernel, iterations=erode_mask)
MASK = MASK.astype(np.bool)
SUN_MASK = data["SUN_MASK"]
#
# In case of forward simulation there are no measurements.
# Use zeros to simplify the implementation of the code.
#
if measurements is not None:
#
# Mask high measurements (sun region).
# Note:
# The masking is done on all channels together
# so that all channels will have the same mask.
#
if mask_measurement > 0:
values_mask = measurements[id_index] < mask_measurement
values_mask = values_mask.min(axis=-1)
MASK = MASK & values_mask
measurement = measurements[id_index][..., color_channel]
else:
measurement = np.zeros_like(PHI)
if subsample > 0:
logging.info(
"Subsampling the measurements by {}".format(subsample)
)
#
# Subsample the sensor smoothly.
#
PHI, PSI, SUN_MASK, measurement = \
subsampleSensor(subsample, PHI, PSI, SUN_MASK, measurement)
#
# The mask is sampled "not" smoothly.
#
MASK = MASK[::subsample, ::subsample]
shape = PHI.shape
self.measurements.append(measurement)
self.shape_array.append(shape)
self.phi_array.append(PHI)
self.psi_array.append(PSI)
self.x_array.append(extra_data["x"]*np.ones(shape=shape)/1000.0)
self.y_array.append(extra_data["y"]*np.ones(shape=shape)/1000.0)
self.z_array.append(extra_data["z"]*np.ones(shape=shape)/1000.0)
#
# Mask any nan values in the images.
#
nan_mask = ~np.isnan(measurement)
self.mask_array.append(MASK & nan_mask)
#
# Scale by view angle.
#
if elevation_weight > 0:
eps = 1e-8
angle_scaling = -np.cos(PSI)
angle_scaling[angle_scaling<eps] = eps
angle_scaling = angle_scaling**elevation_weight
else:
angle_scaling = np.ones_like(PSI)
#
# Weight the scaling by channel.
# This is used for putting more weights on the Blue channel in
# hope to improve the reconstruction of the clouds.
#
angle_scaling = angle_scaling * weights_color[color_channel]
if ignore_sunblock:
self.weights_array.append(angle_scaling)
else:
self.weights_array.append(SUN_MASK*angle_scaling)
#
# SHDOM cant handle negative heights
# move the origin to the lowest camera position
#
minz = np.min(map(lambda x: x.min(), self.z_array))
self.z_array = map(lambda x: x-minz, self.z_array)
#
# Prepare perturbation.
#
if self.perturbate:
self._prepare_camera_perturbation()
#
# Flatten the data arrays.
#
self._flatten_arrays()
/home/addalin/code/pyshdom_C2/scripts/camera_array_cloud_retrieval_RGB.py
take imports from above and sensors, on visl 135, pyshdom_new conda env
# Note:
# I scale the measurements by the SHDOM_COEFF. Theoretically it would
# make sense to apply this to the solarflux input to the SHDOM algorithm.
# This would simplify handling real measurements and simulated images
# in the same way. Practically this is less stable SHDOM wise.
#
measured_images = loadRGBmeasurements(
base_path,
scaling=radiometric_calibration[radiometric_set]
)
/VISL2_net/addalin/efs/SHDOM/exports/2017_05_19/2017_05_19_10_30_00_no_rad_hdr_copyfordbg/
add option to export to dump labeleddata.pkl. Includes: Image, Mask, Sunmask, Sunshader Mask, Cloud weights.
/VISL2_net/addalin/code/pyshdom_C2/shdom/sensors.py
class RGBSensorNetwork(Sensor): """A network of cameras based on the 'real' camera network.
/home/addalin/code/pyshdom_C2/scripts/camera_array_cloud_retrieval_RGB.py take imports from above and sensors, on visl 135, pyshdom_new conda env
radiometric_calibration = dict( may=np.array([0.00243353, 0.00214622, 0.00221308]), sep=np.array([0.00394748, 0.0036557, 0.00418327]), )
#
Load measurements
radiometric_set= 'sep'
EXPORT
/home/shubi/PycharmProjects/cameranetwork/CameraNetwork/export.py