ceccurr's picture
EquiFusion demo
e6d357a verified
Raw History Blame Contribute Delete
4.06 kB
import torch
class Skeleton():
node_hip = {0: 'GlobalRoot'}
node_type = "joint" # graph nodes are joints ('joint') or bones ('bone')
# """
# This class creates a slkeleton that considers hip and pose separately.
# The hip is the first joint, and the pose is the rest of the joints.
# The sequence is centered around the first joint in the sequence, or around the hip joint at frame seq_centering
# It is possible to consider the hip or not.
def __init__(self, if_consider_hip=True, obs_length=50, pred_length=100, seq_centering: int = 0, **kwargs):
self.if_consider_hip = if_consider_hip
self.obs_length = obs_length
self.pred_length = pred_length
assert all([len(val.replace('Hip', ''))>1 for val in self.node_hip.values()]) # in this way it will have a different joint_id mapping from the other centered hip joints.
self.seq_centering = seq_centering
assert self.seq_centering < 0 or self.seq_centering < self.obs_length+self.pred_length, "seq_centering should be less than total seq length"
################# functions to obtain input poses ########################################################################################
# THese funcitons take all kpts as input and return all kpts as output (not only pose or hip)
def _get_where_is_seq_centered(self):
if self.seq_centering < 0:
return self.obs_length+self.seq_centering
assert 0, "This was changed from - to +, so your old exps are not reproducible..."
else:
return self.seq_centering
def transform_hip_to_input_space(self, data):
centered, hips = data[..., 1:, :], data[..., 0:1, :]
#set beginning of seq to where_isZero
where_isZero = self._get_where_is_seq_centered()
hips = hips - hips[...,where_isZero,:,:].unsqueeze(-3)
return torch.cat([hips, centered], dim=-2)
def tranform_to_input_space(self, data):
data = self.tranform_to_input_space_pose_only(data)
# centered, hips = data[..., 1:, :], data[..., 0, :].unsqueeze(-2)
if not self.if_consider_hip:
kpts = data[..., 1:, :]
else:
kpts = self.transform_hip_to_input_space(data)
return kpts
def tranform_to_input_space_pose_only(self, data):
...
return data
def add_zero_pad_center_hip(self, kpts):
shape = list(kpts.shape)
shape[-2] = 1
kpts = torch.cat([torch.zeros(shape, device=kpts.device), kpts], dim=-2)
return kpts
def if_add_zero_pad_center_hip(self, kpts):
if not self.if_consider_hip and kpts.shape[-2] == self.num_joints - 1:
kpts =self.add_zero_pad_center_hip(kpts)
return kpts
########################## function to tranform back to metric space ##########################################################################
def transform_hip_to_metric_space(self, kpts):
...
return kpts
def _merge_hip_and_poseinmetricspace(self, hip_coords, kpts):
...
return torch.cat([hip_coords, kpts], dim=-2)
def transform_to_metric_space(self, kpts):
"""_summary_
Returns:
kpts: kpts in 3D coordinates
"""
if self.if_consider_hip:
kpts = self.transform_hip_to_metric_space(kpts) # hip is now represented in metric space, so it is first location.
hip_coords = kpts[...,:1,:]
kpts = self.transform_to_metric_space_pose_only(kpts[...,1:,:])
return self._merge_hip_and_poseinmetricspace(hip_coords, kpts)
else:
# input is without hips.
kpts = self.transform_to_metric_space_pose_only(kpts)
return kpts
def transform_to_metric_space_pose_only(self, kpts):
"""This is the only funtion that tkaes as input the pose and returns the pose joints"""
...
return kpts