def __call__(self, results):
# stack=False
for key in [
'points', 'pts_indices', 'seg_points', 'seg_pts_indices',
'seg_label', 'scn_coords'
]:
# assert key in results.keys(), f"key {key} does not exist."
if key not in results.keys():
continue
results[key] = DC(to_tensor(results[key]), stack=False)
# stack=True
img = np.ascontiguousarray(results['img'].transpose(2, 0, 1))
results['img'] = DC(to_tensor(img), stack=True)
# gt_labels_3d: list of int
for key in ['gt_labels_3d']:
if key not in results:
continue
if isinstance(results[key], list):
results[key] = DC([to_tensor(res) for res in results[key]])
else:
results[key] = DC(to_tensor(results[key]))
if 'gt_bboxes_3d' in results:
if isinstance(results['gt_bboxes_3d'], BaseInstance3DBoxes):
results['gt_bboxes_3d'] = DC(results['gt_bboxes_3d'],
cpu_only=True)
else:
results['gt_bboxes_3d'] = DC(to_tensor(
results['gt_bboxes_3d']))
return results
def reid_format_bundle(self, results):
"""Transform and format gt_label fields in results.
Args:
results (dict): Result dict contains the data to convert.
Returns:
dict: The result dict contains the data that is formatted with
ReID bundle.
"""
for key in results:
if key == 'img':
img = results[key]
if img.ndim == 3:
img = np.ascontiguousarray(img.transpose(2, 0, 1))
else:
img = np.ascontiguousarray(img.transpose(3, 2, 0, 1))
results['img'] = DC(to_tensor(img), stack=True)
elif key == 'gt_label':
results[key] = DC(to_tensor(results[key]),
stack=True,
pad_dims=None)
else:
raise KeyError(f'key {key} is not supported')
return results
def __call__(self, results):
if 'img' in results:
img = np.ascontiguousarray(results['img'].transpose(2, 0, 1))
results['img'] = DC(to_tensor(img).float(), stack=True)
if 'gt_masks' in results:
results['gt_masks'] = DC(to_tensor(results['gt_masks']),
stack=True)
return results
def __call__(self, results):
if 'img' in results:
img = np.ascontiguousarray(results['img'].transpose(2, 0, 1))
results['img'] = DC(to_tensor(img), stack=True)
for key in ['proposals', 'gt_bboxes', 'gt_coefs', 'gt_skeleton', 'gt_bboxes_ignore', 'gt_labels']:
if key not in results:
continue
results[key] = DC(to_tensor(results[key]))
if 'gt_masks' in results:
results['gt_masks'] = DC(results['gt_masks'], cpu_only=True)
if 'gt_semantic_seg' in results:
results['gt_semantic_seg'] = DC(
to_tensor(results['gt_semantic_seg'][None, ...]), stack=True)
return results
def images_to_tensor(self, results):
"""Transpose and convert images/multi-images to Tensor."""
if 'img' in results:
img = results['img']
if len(img.shape) == 3:
# (H, W, 3) to (3, H, W)
img = np.ascontiguousarray(img.transpose(2, 0, 1))
else:
# (H, W, 3, N) to (N, 3, H, W)
img = np.ascontiguousarray(img.transpose(3, 2, 0, 1))
results['img'] = to_tensor(img)
if 'proposals' in results:
results['proposals'] = to_tensor(results['proposals'])
if 'img_metas' in results:
results['img_metas'] = DC(results['img_metas'], cpu_only=True)
return results
def __call__(self, results):
"""Call function to transform and format common fields in results.
Args:
results (dict): Result dict contains the data to convert.
Returns:
dict: The result dict contains the data that is formatted with
default bundle.
"""
# Format 3D data
if 'points' in results:
assert isinstance(results['points'], BasePoints)
results['points'] = DC(results['points'].tensor)
for key in ['voxels', 'coors', 'voxel_centers', 'num_points']:
if key not in results:
continue
results[key] = DC(to_tensor(results[key]), stack=False)
if self.with_gt:
# Clean GT bboxes in the final
if 'gt_bboxes_3d_mask' in results:
gt_bboxes_3d_mask = results['gt_bboxes_3d_mask']
results['gt_bboxes_3d'] = results['gt_bboxes_3d'][
gt_bboxes_3d_mask]
if 'gt_names_3d' in results:
results['gt_names_3d'] = results['gt_names_3d'][
gt_bboxes_3d_mask]
if 'gt_bboxes_mask' in results:
gt_bboxes_mask = results['gt_bboxes_mask']
if 'gt_bboxes' in results:
results['gt_bboxes'] = results['gt_bboxes'][gt_bboxes_mask]
results['gt_names'] = results['gt_names'][gt_bboxes_mask]
if self.with_label:
if 'gt_names' in results and len(results['gt_names']) == 0:
results['gt_labels'] = np.array([], dtype=np.int64)
elif 'gt_names' in results and isinstance(
results['gt_names'][0], list):
# gt_labels might be a list of list in multi-view setting
results['gt_labels'] = [
np.array([self.class_names.index(n) for n in res],
dtype=np.int64) for res in results['gt_names']
]
elif 'gt_names' in results:
results['gt_labels'] = np.array([
self.class_names.index(n) for n in results['gt_names']
],
dtype=np.int64)
# we still assume one pipeline for one frame LiDAR
# thus, the 3D name is list[string]
if 'gt_names_3d' in results:
results['gt_labels_3d'] = np.array([
self.class_names.index(n)
for n in results['gt_names_3d']
],
dtype=np.int64)
results = super(DefaultFormatBundle3D, self).__call__(results)
return results
def __call__(self, results):
outs = []
for _results in results:
_results = super().__call__(_results)
_results['gt_match_indices'] = DC(
to_tensor(_results['gt_match_indices']))
outs.append(_results)
return outs
def default_format_bundle(self, results):
"""Transform and format common fields in results.
Args:
results (dict): Result dict contains the data to convert.
Returns:
dict: The result dict contains the data that is formatted with
default bundle.
"""
if 'img' in results:
img = results['img']
if len(img.shape) == 3:
img = np.ascontiguousarray(img.transpose(2, 0, 1))
else:
img = np.ascontiguousarray(img.transpose(3, 2, 0, 1))
results['img'] = DC(to_tensor(img), stack=True)
if 'padding_mask' in results:
results['padding_mask'] = DC(
to_tensor(results['padding_mask'].copy()), stack=True)
for key in [
'proposals', 'gt_bboxes', 'gt_bboxes_ignore', 'gt_labels',
'gt_instance_ids', 'gt_match_indices'
]:
if key not in results:
continue
results[key] = DC(to_tensor(results[key]))
for key in ['img_metas', 'gt_masks']:
if key in results:
results[key] = DC(results[key], cpu_only=True)
if 'gt_semantic_seg' in results:
semantic_seg = results['gt_semantic_seg']
if len(semantic_seg.shape) == 2:
semantic_seg = semantic_seg[None, ...]
else:
semantic_seg = np.ascontiguousarray(
semantic_seg.transpose(3, 2, 0, 1))
results['gt_semantic_seg'] = DC(
to_tensor(results['gt_semantic_seg']), stack=True)
return results
def prepare_train_img(self, idx):
""" Pipelines same as PairDET30Datase. """
img_info = self.img_infos[idx]
ann_info = self.get_ann_info(idx)
results = dict(img_info=img_info, ann_info=ann_info)
self.pre_pipeline(results)
results = self.pipeline(results)
results['ref_img'] = DC(results['img'].data.clone(), stack=True)
results['ref_img_meta'] = results['img_meta'] # Is it read-only ?
results['ref_bboxes'] = DC(results['gt_bboxes'].data.clone())
results['ref_labels'] = DC(results['gt_labels'].data.clone())
# We generate 'trackids' field manually
num_gts = len(results['gt_labels'].data)
_data = [i + 1 for i in range(num_gts)
] # note we leave 0 for negative/background
trackids = results['gt_labels'].data.new_tensor(_data)
results['gt_trackids'] = DC(to_tensor(trackids))
results['ref_trackids'] = DC(to_tensor(trackids))
if len(results['gt_bboxes'].data) == 0:
return None
return results
def __call__(self, results):
"""Call function to transform and format common fields in results.
Args:
results (dict): Result dict contains the data to convert.
Returns:
dict: The result dict contains the data that is formatted with
default bundle.
"""
if 'img' in results:
if isinstance(results['img'], list):
# process multiple imgs in single frame
imgs = [img.transpose(2, 0, 1) for img in results['img']]
imgs = np.ascontiguousarray(np.stack(imgs, axis=0))
results['img'] = DC(to_tensor(imgs), stack=True)
else:
img = np.ascontiguousarray(results['img'].transpose(2, 0, 1))
results['img'] = DC(to_tensor(img), stack=True)
for key in [
'proposals', 'gt_bboxes', 'gt_bboxes_ignore', 'gt_labels',
'gt_labels_3d', 'attr_labels', 'pts_instance_mask',
'pts_semantic_mask', 'centers2d', 'depths'
]:
if key not in results:
continue
if isinstance(results[key], list):
results[key] = DC([to_tensor(res) for res in results[key]])
else:
results[key] = DC(to_tensor(results[key]))
if 'gt_bboxes_3d' in results:
if isinstance(results['gt_bboxes_3d'], BaseInstance3DBoxes):
results['gt_bboxes_3d'] = DC(
results['gt_bboxes_3d'], cpu_only=True)
else:
results['gt_bboxes_3d'] = DC(
to_tensor(results['gt_bboxes_3d']))
if 'gt_masks' in results:
results['gt_masks'] = DC(results['gt_masks'], cpu_only=True)
if 'gt_semantic_seg' in results:
results['gt_semantic_seg'] = DC(
to_tensor(results['gt_semantic_seg'][None, ...]), stack=True)
if 'bev_seg_image' in results:
results['bev_seg_image'] = DC(to_tensor(results['bev_seg_image'][None, ...]), stack=True)
return results
def __call__(self, results):
"""Call function to transform and format common fields in results.
Args:
results (dict): Result dict contains the data to convert.
Returns:
dict: The result dict contains the data that is formatted with
default bundle.
"""
# Format 3D data
if 'points' in results:
assert isinstance(results['points'], BasePoints)
results['points'] = DC(results['points'].tensor)
for key in ['voxels', 'coors', 'voxel_centers', 'num_points']:
if key not in results:
continue
results[key] = DC(to_tensor(results[key]), stack=False)
# Feng Xiang begin
# code begin
# print("FENGXIANGRESULTSKEYS:LSDKJF:LKSDJF:LSDKJF")
# print("Self With GT: " + str(self.with_gt)) # TRUE
# print("Self with Label: " + str(self.with_label)) # FALSE
# print(results.keys())
# code end
if self.with_gt:
# Clean GT bboxes in the final
if 'gt_bboxes_3d_mask' in results:
print("Formating - DefaultFormatBundle3D - gt_bboxes_3d_mask")
gt_bboxes_3d_mask = results['gt_bboxes_3d_mask']
results['gt_bboxes_3d'] = results['gt_bboxes_3d'][
gt_bboxes_3d_mask]
if 'gt_names_3d' in results:
results['gt_names_3d'] = results['gt_names_3d'][
gt_bboxes_3d_mask]
# Feng Xiang code
# code begin
results['gt_attr_3d'] = results['gt_attr_3d'][gt_bboxes_3d_mask]
# code end
if 'centers2d' in results:
results['centers2d'] = results['centers2d'][
gt_bboxes_3d_mask]
if 'depths' in results:
results['depths'] = results['depths'][gt_bboxes_3d_mask]
if 'gt_bboxes_mask' in results:
print("Formatting - DefaultFormatBundle3D - gt_bboxes_mask")
gt_bboxes_mask = results['gt_bboxes_mask']
if 'gt_bboxes' in results:
results['gt_bboxes'] = results['gt_bboxes'][gt_bboxes_mask]
results['gt_names'] = results['gt_names'][gt_bboxes_mask]
# Feng Xiang code
# code begin
results['gt_attr'] = results['gt_attr'][gt_bboxes_mask]
# code end
if self.with_label:
# print("Formatting - DefaultFormatBundle3D - self.with_label")
if 'gt_names' in results and len(results['gt_names']) == 0:
print("Formatting - DefaultFormatBundle3D - self.with_label - gt_names and len")
results['gt_labels'] = np.array([], dtype=np.int64)
results['attr_labels'] = np.array([], dtype=np.int64)
# Feng Xiang code
# code begin
results['gt_attr'] = np.array([], dtype=np.int64)
# code end
elif 'gt_names' in results and isinstance(
results['gt_names'][0], list):
print("Formatting - DefaultFormatBundle3D - self.with_label - gt_names and isinstance")
# gt_labels might be a list of list in multi-view setting
results['gt_labels'] = [
np.array([self.class_names.index(n) for n in res],
dtype=np.int64) for res in results['gt_names']
]
# Feng Xiang code
# code begin
results['gt_attr'] = [np.array([self.attr_names.index(n) for n in res], dtype=np.int64) for res in results['gt_names']]
# code end
elif 'gt_names' in results:
print("Formatting - DefaultFormatBundle3D - self.with_label - gt_names only")
results['gt_labels'] = np.array([
self.class_names.index(n) for n in results['gt_names']
], dtype=np.int64)
results['gt_attr'] = np.array([
self.attr_names.index(n) for n in results['gt_names']
], dtype=np.int64)
# we still assume one pipeline for one frame LiDAR
# thus, the 3D name is list[string]
if 'gt_names_3d' in results:
print("Formatting - DefaultFormatBundle3D - gt_names_3d")
results['gt_labels_3d'] = np.array([
self.class_names.index(n)
for n in results['gt_names_3d']
], dtype=np.int64)
results['gt_attr_3d'] = np.array([
self.attr_names.index(n)
for n in results['gt_names_3d']
], dtype=np.int64)
results = super(DefaultFormatBundle3D, self).__call__(results)
return results
def __call__(self, results):
for key in self.keys:
results[key] = to_tensor(
np.expand_dims(results[key].transpose(2, 0, 1), 0))
return results
