Learned bone lengths and relative marker positions as well as pose variables (i.e. bone rotations and translations) for all time points of the sequence, which is used for learning the animal's anatomy. The format is:
np.ndarray, shape=(nBones + 3*nMarkers + (3*nBones+3)*nFrames_calib,), dtype('float64'),
Thus, the individual variables are stored in x_calib.npy according to:
x_calib[0 : nBones] # bone lengths
x_calib[nBones : nBones+3*nMarkers] # relative 3D marker positions
x_calib[nBones+3*nMarkers : nBones+3*nMarkers+(3*nBones+3)*1] # bone rotations and translation for time point 1
x_calib[nBones+3*nMarkers+(3*nBones+3)*1 : nBones+3*nMarkers+(3*nBones+3)*2] # bone rotations and translation for time point 2
x_calib[nBones+3*nMarkers+(3*nBones+3)*2 : nBones+3*nMarkers+(3*nBones+3)*3] # bone rotations and translation for time point 3
.
.
.
x_calib[nBones+3*nMarkers+(3*nBones+3)*(nFrames_calib-1) : nBones+3*nMarkers+(3*nBones+3)*nFrames_calib] # bone rotations and translation for time point nFrames_calib
Note that bone lengths entries for right-sided bones are always zero, as the actual values are copied from the corresponding left-sided bone lengths entries to enforce symmetry.
Learned bone lengths and relative marker positions as well as pose variables (i.e. bone rotations and a single translation) for the first time point of the sequence, which should be reconstructed. The format is:
np.ndarray, shape=(nBones + 3*nMarkers + 3*nBones+3,), dtype('float64'),
Thus, the individual variables are stored in x_ini.npy according to:
x_ini[0 : nBones] # bone lengths
x_ini[nBones : nBones+3*nMarkers] # relative 3D marker positions
x_ini[nBones+3*nMarkers : nBones+3*nMarkers+(3*nBones+3)] # bone rotations and translation
Thus, the first nBones+3*nMarkers entries of x_ini.npy are identical to x_calib.npy.
Learned model parameters as well as resulting latent variables and corresponding covariance matrices for the entire sequence. If temporal constraints are enforced (i.e. mode=4 or mode=3 in calibration.py), the format is:
dict(
# learned transition matrix (legacy: this is fixed to the identity matrix) # LEGACY
'A': np.ndarray, shape=(nLatent, nLatent,), dtype('float64'), # LEGACY
# learned inital state of the latent variables
'mu0': np.ndarray, shape=(nLatent,), dtype('float64'),
# inferred latent variables (inferrence is based on the learned model parameters)
'mu_uks': np.ndarray, shape=(nFrames+1, nLatent,), dtype('float64'),
# learned initial covariance matrix of the latent variables
'var0': np.ndarray, shape=(nLatent, nLatent,), dtype('float64'),
# learned covariance matrix of the transition noise
'var_f': np.ndarray, shape=(nLatent, nLatent,), dtype('float64'),
# learned covariance matrix of the measurement noise
'var_g': np.ndarray, shape=(nMeasurement, nMeasurement,), dtype('float64'),
# inferred covariance matrices of the latent variables (inferrence is based on the learned model parameters)
'var_uks': np.ndarray, shape=(nFrames+1, nLatent, nLatent,), dtype('float64'),
)
If only anatomical or no constraints are enforced (i.e. mode=2 or mode=1 in calibration.py), the format is:
dict(
# inferred latent variables (inferrence is based on the learned model parameters; mu_fit[0] is filled with nan's, since it corresponds to mu0)
'mu_fit': np.ndarray, shape=(nFrames+1, nLatent,), dtype('float64'),
)
Inferred 3D joint and marker positions as well as resulting projected marker locations in the 2D images. The format is:
dict(
# reconstructed marker locations in 2D image
'marker_positions_2d': np.ndarray, shape=(nFrames, nCameras, nMarkers, 2,), dtype('float64'),
# reconstructed marker positions in 3D
'marker_positions_3d': np.ndarray, shape=(nFrames, nMarkers, 3,), dtype('float64'),
# reconstructed joint positions in 3D
'joint_positions_3d': np.ndarray, shape=(nFrames, nJoints, 3,), dtype('float64'),
)