def use_filter(x):
structure = skeletalModel.getSkeletalModelStructure()
inputSequence_2D = x
# Decomposition of the single matrix into three matrices: x, y, w (=likelihood)
X = inputSequence_2D
Xx = X[0:X.shape[0], 0:(X.shape[1]):3]
Xy = X[0:X.shape[0], 1:(X.shape[1]):3]
Xw = X[0:X.shape[0], 2:(X.shape[1]):3]
# Normalization of the picture (x and y axis has the same scale)
Xx, Xy = pose2D.normalization(Xx, Xy)
# Delete all skeletal models which have a lot of missing parts.
Xx, Xy, Xw = pose2D.prune(Xx, Xy, Xw, (0, 1, 2, 3, 4, 5, 6, 7), 0.3, dtype)
# Preliminary filtering: weighted linear interpolation of missing points.
Xx, Xy, Xw = pose2D.interpolation(Xx, Xy, Xw, 0.99, dtype)
# Initial 3D pose estimation
lines0, rootsx0, rootsy0, rootsz0, anglesx0, anglesy0, anglesz0, Yx0, Yy0, Yz0 = pose2Dto3D.initialization(
Xx,
Xy,
Xw,
structure,
0.001, # weight for adding noise
randomNubersGenerator,
dtype,
percentil=0.5,
)
# Backpropagation-based filtering
Yx, Yy, Yz = pose3D.backpropagationBasedFiltering(
lines0,
rootsx0,
rootsy0,
rootsz0,
anglesx0,
anglesy0,
anglesz0,
Xx,
Xy,
Xw,
structure,
dtype,
nCycles=100,
)
return convList2Array([Yx, Yy, Yz])
Xx, Xy, Xw = pose2D.interpolation(Xx, Xy, Xw, 0.99, dtype)
save("data/demo3.txt", [Xx, Xy, Xw])
# Initial 3D pose estimation
lines0, rootsx0, rootsy0, rootsz0, anglesx0, anglesy0, anglesz0, Yx0, Yy0, Yz0 = pose2Dto3D.initialization(
Xx,
Xy,
Xw,
structure,
0.001, # weight for adding noise
randomNubersGenerator,
dtype)
save("data/demo4.txt", [Yx0, Yy0, Yz0])
# Backpropagation-based filtering
Yx, Yy, Yz = pose3D.backpropagationBasedFiltering(
lines0,
rootsx0,
rootsy0,
rootsz0,
anglesx0,
anglesy0,
anglesz0,
Xx,
Xy,
Xw,
structure,
dtype,
)
save("data/demo5.txt", [Yx, Yy, Yz])
def main(argv):
dtype = "float32"
randomNubersGenerator = numpy.random.RandomState(1234)
# This demo shows converting a result of 2D pose estimation into a 3D pose.
# Getting our structure of skeletal model.
# For customizing the structure see a definition of getSkeletalModelStructure.
structure = skeletalModel.getSkeletalModelStructure()
# Getting 2D data
# The sequence is an N-tuple of
# (1sf point - x, 1st point - y, 1st point - likelihood, 2nd point - x, ...)
# a missing point should have x=0, y=0, likelihood=0
f = h5py.File("./pipelineDemo/keypoints.h5", "r")
# todo change the location or folder name later
inputSequence_2D = numpy.array(f.get(argv[1]))
#inputSequence_2D = numpy.array(f.get("sample_2d_openpose"))
f.close()
# Decomposition of the single matrix into three matrices: x, y, w (=likelihood)
X = inputSequence_2D
Xx = X[0:X.shape[0], 0:150:3]
Xy = X[0:X.shape[0], 1:150:3]
Xw = X[0:X.shape[0], 2:150:3]
Face_Xx = X[0:X.shape[0], 150::3]
Face_Xy = X[0:X.shape[0], 150::3]
Face_Xw = X[0:X.shape[0], 150::3]
# Normalization of the picture (x and y axis has the same scale)
Xx, Xy = pose2D.normalization(Xx, Xy)
Face_Xx, Face_Xy = pose2D.normalization(Face_Xx, Face_Xy)
save("Temp_Data/" + argv[1] + "_demo1.txt", [Xx, Xy, Xw])
# Delete all skeletal models which have a lot of missing parts.
Xx, Xy, Xw = pose2D.prune(Xx, Xy, Xw, (0, 1, 2, 3, 4, 5, 6, 7), 0.3, dtype)
save("Temp_Data/" + argv[1] + "_demo2.txt", [Xx, Xy, Xw])
# Preliminary filtering: weighted linear interpolation of missing points.
Xx, Xy, Xw = pose2D.interpolation(Xx, Xy, Xw, 0.99, dtype)
save("Temp_Data/" + argv[1] + "_demo3.txt", [Xx, Xy, Xw])
# Initial 3D pose estimation
lines0, rootsx0, rootsy0, rootsz0, anglesx0, anglesy0, anglesz0, Yx0, Yy0, Yz0 = pose2Dto3D.initialization(
Xx,
Xy,
Xw,
structure,
0.001, # weight for adding noise
randomNubersGenerator,
dtype)
save("Temp_Data/" + argv[1] + "_demo4.txt", [Yx0, Yy0, Yz0])
# Backpropagation-based filtering
Yx, Yy, Yz = pose3D.backpropagationBasedFiltering(
lines0,
rootsx0,
rootsy0,
rootsz0,
anglesx0,
anglesy0,
anglesz0,
Xx,
Xy,
Xw,
structure,
dtype,
)
data = numpy.asarray(noNones([
numpy.hstack((Yx, Face_Xx)),
numpy.hstack((Yy, Face_Xy)),
numpy.hstack((Yz, Face_Xw))
]),
dtype="float32")
save("Training_Data_3D/" + argv[1] + ".txt", data)
hf = h5py.File("Training_Data_3D_h5/" + argv[1] + ".h5", "w")
hf.create_dataset(argv[1], data=data, dtype="float32")
hf.close()