423A35C7/pytorch3d
1
0
Fork 0
mirror of https://github.com/facebookresearch/pytorch3d.git synced 2025-12-20 06:10:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

initialize pointcloud from list containing Nones

Browse Source 
Summary:
The following snippet should work in more cases.
     point_cloud = Pointclouds(
         [pcl.points_packed() for pcl in point_clouds],
         features=[pcl.features_packed() for pcl in point_clouds],
     )

We therefore allow features and normals inputs to be lists which contain some (but not all) Nones.

The initialization of a Pointclouds from empty data is also made a bit better now at working out how many feature channels there are.

Reviewed By: davnov134

Differential Revision: D31795089

fbshipit-source-id: 54bf941ba80672d699ffd5ac28927740e830f8ab
This commit is contained in:
Jeremy Reizenstein
2022-01-07 05:50:54 -08:00
committed by Facebook GitHub Bot
parent 9640560541
commit fc4dd80208
2 changed files with 105 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

92
pytorch3d/structures/pointclouds.py
View File

@@ -266,30 +266,8 @@ class Pointclouds:
aux_input_C = None
if isinstance(aux_input, list):
if len(aux_input) != self._N:
raise ValueError("Points and auxiliary input must be the same length.")
for p, d in zip(self._num_points_per_cloud, aux_input):
if p != d.shape[0]:
raise ValueError(
"A cloud has mismatched numbers of points and inputs"
)
if d.device != self.device:
raise ValueError(
"All auxiliary inputs must be on the same device as the points."
)
if p > 0:
if d.dim() != 2:
raise ValueError(
"A cloud auxiliary input must be of shape PxC or empty"
)
if aux_input_C is None:
aux_input_C = d.shape[1]
if aux_input_C != d.shape[1]:
raise ValueError(
"The clouds must have the same number of channels"
)
return aux_input, None, aux_input_C
elif torch.is_tensor(aux_input):
return self._parse_auxiliary_input_list(aux_input)
if torch.is_tensor(aux_input):
if aux_input.dim() != 3:
raise ValueError("Auxiliary input tensor has incorrect dimensions.")
if self._N != aux_input.shape[0]:
@@ -312,6 +290,72 @@ class Pointclouds:
points in a cloud."
)
def _parse_auxiliary_input_list(
self, aux_input: list
) -> Tuple[Optional[List[torch.Tensor]], None, Optional[int]]:
"""
Interpret the auxiliary inputs (normals, features) given to __init__,
if a list.
Args:
aux_input:
- List where each element is a tensor of shape (num_points, C)
containing the features for the points in the cloud.
For normals, C = 3
Returns:
3-element tuple of list, padded=None, num_channels.
If aux_input is list, then padded is None. If aux_input is a tensor,
then list is None.
"""
aux_input_C = None
good_empty = None
needs_fixing = False
if len(aux_input) != self._N:
raise ValueError("Points and auxiliary input must be the same length.")
for p, d in zip(self._num_points_per_cloud, aux_input):
valid_but_empty = p == 0 and d is not None and d.ndim == 2
if p > 0 or valid_but_empty:
if p != d.shape[0]:
raise ValueError(
"A cloud has mismatched numbers of points and inputs"
)
if d.dim() != 2:
raise ValueError(
"A cloud auxiliary input must be of shape PxC or empty"
)
if aux_input_C is None:
aux_input_C = d.shape[1]
elif aux_input_C != d.shape[1]:
raise ValueError("The clouds must have the same number of channels")
if d.device != self.device:
raise ValueError(
"All auxiliary inputs must be on the same device as the points."
)
else:
needs_fixing = True
if aux_input_C is None:
# We found nothing useful
return None, None, None
# If we have empty but "wrong" inputs we want to store "fixed" versions.
if needs_fixing:
if good_empty is None:
good_empty = torch.zeros((0, aux_input_C), device=self.device)
aux_input_out = []
for p, d in zip(self._num_points_per_cloud, aux_input):
valid_but_empty = p == 0 and d is not None and d.ndim == 2
if p > 0 or valid_but_empty:
aux_input_out.append(d)
else:
aux_input_out.append(good_empty)
else:
aux_input_out = aux_input
return aux_input_out, None, aux_input_C
def __len__(self) -> int:
return self._N