def abs_to_hiprel(poses, joint_set):
""" Converts an absolute pose into [hi]+relative_pose. """
assert_shape(poses, (None, joint_set.NUM_JOINTS, 3))
if isinstance(poses, torch.Tensor):
root = poses[:, [joint_set.index_of('hip')]]
rel = remove_root(poses, joint_set.index_of('hip'))
return torch.cat([root, rel], dim=-2)
else:
root = poses[:, [joint_set.index_of('hip')]].copy()
rel = remove_root(poses, joint_set.index_of('hip'))
return np.concatenate([root, rel], axis=-2)
def eval(self, model=None):
losses, preds = self.pred_and_calc_loss(model)
losses = np.concatenate([losses[seq] for seq in self.sequences])
self.val_loss = np.nanmean(losses)
self.losses_to_log = {self.prefix + '_loss': self.val_loss}
self.losses = losses
self.preds = preds
# Assuming hip is the last component
if self.is_absolute:
self.losses_to_log[self.prefix + '_abs_loss'] = np.nanmean(
losses[:, -3:])
self.losses_to_log[self.prefix + '_rel_loss'] = np.nanmean(
losses[:, :-3])
else:
self.losses_to_log[self.prefix + '_rel_loss'] = self.val_loss
assert self.pctiles[
-1] == 99, "Currently the last percentile is hardcoded to be 99 for printing"
sequence_mpjpes, sequence_pcks, sequence_pctiles, joint_means, joint_pctiles = \
eval_results(preds, self.data_3d_mm, self.joint_set, pctiles=self.pctiles, verbose=True)
# Calculate relative error
if self.is_absolute:
rel_pred = {}
rel_gt = {}
for seq in preds:
rel_pred[seq] = remove_root(preds[seq],
self.joint_set.index_of('hip'))
rel_gt[seq] = remove_root(self.data_3d_mm[seq],
self.joint_set.index_of('hip'))
rel_mean_error, _, _, _, _ = eval_results(rel_pred,
rel_gt,
self.joint_set,
verbose=False)
rel_mean_error = np.mean(list(rel_mean_error.values()))
print("Root relative error: %.2f mm" % rel_mean_error)
self.rel_mean_error = rel_mean_error
self.losses_to_log[self.prefix + '_rel_error'] = rel_mean_error
self.mean_sequence_mpjpe = np.mean(list(sequence_mpjpes.values()))
self.mean_sequence_pck = np.mean(list(sequence_pcks.values()))
self.losses_to_log[self.prefix + '_err'] = self.mean_sequence_mpjpe
self.losses_to_log[self.prefix + '_pck'] = self.mean_sequence_pck
return sequence_mpjpes, sequence_pcks, sequence_pctiles, joint_means, joint_pctiles
def preprocess_3d(data, add_root, log_root_z, joint_set, root_name):
"""
3D preprocessing:
1. Removes the root joint
2. If add_root is True, append the root joint at the end of the pose. The
The logarithm of the z coordinate of the root is taken.
3. Flattens the data.
:param data: ndarray(nFrames, [nPoses], nJoints, 3[x, y, z]) 3D coordinates in MuPoTS order
:param add_root: True if the absolute coordinates of the hip should be included in the output
:param log_root_z:if true, the log of the z coordinate of the root is used
:param root_name: name of the root joint, must be a MuPoTS joint
:return: ndarray(nPoses, 3*nJoints|3*(nJoints-1)), 3*nJoints if add_root is true otherwise 3*(nJoints-1)
"""
assert_shape(data, ("*", joint_set.NUM_JOINTS, 3))
root_ind = joint_set.index_of(root_name)
root3d = data[..., root_ind, :].copy()
if log_root_z:
root3d[..., 2] = np.log(root3d[..., 2])
data = remove_root(data, root_ind) # (nFrames, [nPoses], nJoints-1, 3)
data = data.reshape(data.shape[:-2] + (-1,)) # (nFrames, [nPoses], (nJoints-1)*3)
if add_root:
data = np.concatenate([data, root3d], axis=-1) # (nFrames, [nPoses], nJoints*3)
return data.astype('float32')
def abs_to_hiprel(poses, joint_set):
""" Converts an absolute pose into [hi]+relative_pose. """
assert_shape(poses, (None, joint_set.NUM_JOINTS, 3))
root = poses[:, [joint_set.index_of('hip')]].copy()
rel = remove_root(poses, joint_set.index_of('hip'))
return np.concatenate([root, rel], axis=-2)
def eval(self, model=None, calculate_scale_free=False, verbose=False):
"""
:param model: the evaluator can use this model, if self.model is nor provided
:param calculate_scale_free: if True, also calculates N-MRPE and N_RMPJPE
:return:
"""
losses, preds = self.pred_and_calc_loss(model)
losses = np.concatenate([losses[seq] for seq in self.sequences])
self.val_loss = np.nanmean(losses)
self.losses_to_log = {self.prefix + '_loss': self.val_loss}
self.losses = losses
self.preds = preds
# Assuming hip is the last component
if self.is_absolute:
self.losses_to_log[self.prefix + '_abs_loss'] = np.nanmean(
losses[:, -3:])
self.losses_to_log[self.prefix + '_rel_loss'] = np.nanmean(
losses[:, :-3])
else:
self.losses_to_log[self.prefix + '_rel_loss'] = self.val_loss
assert self.pctiles[
-1] == 99, "Currently the last percentile is hardcoded to be 99 for printing"
sequence_mpjpes, sequence_pcks, sequence_pctiles, joint_means, joint_pctiles = \
eval_results(preds, self.data_3d_mm, self.joint_set, pctiles=self.pctiles, verbose=verbose)
self.losses_to_log[self.prefix + '_mrpe'] = np.mean([
mrpe(preds[s], self.data_3d_mm[s], self.joint_set) for s in preds
])
# Calculate relative error
if self.is_absolute:
rel_pred = {}
rel_gt = {}
for seq in preds:
rel_pred[seq] = remove_root(preds[seq],
self.joint_set.index_of('hip'))
rel_gt[seq] = remove_root(self.data_3d_mm[seq],
self.joint_set.index_of('hip'))
rel_mean_error, _, _, _, _ = eval_results(rel_pred,
rel_gt,
self.joint_set,
verbose=False)
rel_mean_error = np.mean(
np.asarray(list(rel_mean_error.values()), dtype=np.float32))
if verbose:
print("Root relative error (MPJPE): %.2f" % rel_mean_error)
self.rel_mean_error = rel_mean_error
self.losses_to_log[self.prefix + '_rel_error'] = rel_mean_error
self.mean_sequence_mpjpe = np.mean(
np.asarray(list(sequence_mpjpes.values()), dtype=np.float32))
self.mean_sequence_pck = np.mean(
np.asarray(list(sequence_pcks.values()), dtype=np.float32))
self.losses_to_log[self.prefix + '_err'] = self.mean_sequence_mpjpe
self.losses_to_log[self.prefix + '_pck'] = self.mean_sequence_pck
if calculate_scale_free:
scaled_preds = {}
for seq in preds:
# predict a single scale for the full video
pred_points = preds[seq].reshape(1, -1, 3)
gt_points = self.data_3d_mm[seq].reshape(1, -1, 3)
s = optimal_scaling(pred_points, gt_points)
scaled_preds[seq] = preds[seq] * s
n_mrpe = np.mean([
mrpe(scaled_preds[s], self.data_3d_mm[s], self.joint_set)
for s in scaled_preds
])
n_rmpjpe = np.mean([
r_mpjpe(scaled_preds[s], self.data_3d_mm[s], self.joint_set)
for s in scaled_preds
])
if verbose:
print('N-MRPE: %.1f' % n_mrpe)
print('N-MPJPE: %.1f' % n_rmpjpe)
self.losses_to_log[self.prefix + '_n_mrpe'] = n_mrpe
self.losses_to_log[self.prefix + '_n_rel_err'] = n_rmpjpe
return sequence_mpjpes, sequence_pcks, sequence_pctiles, joint_means, joint_pctiles