def _merge_data(batch):
# C x H x W
max_shape = tuple(
max(img.shape[d] for img, _ in batch) for d in range(3))
image_batch = []
labels_batch = None
if is_multi_label:
labels_batch = []
for image_idx, (img, anns) in enumerate(batch):
temp = torch.zeros(max_shape, dtype=torch.float32)
temp[:, :img.shape[1], :img.shape[2]] = img
temp = normalize(temp)
image_batch.append(temp)
if is_multi_label:
labels_batch.append(train_sets.handle_multi_label(anns))
image_batch = torch.stack(image_batch, 0)
anchors = generate_shifted_anchors(max_shape[1:])
rgr, cls = anchor_targets_bbox(anchors,
batch,
len(train_sets.class2idx),
multi_labels=labels_batch)
reg = torch.from_numpy(rgr)
cls = torch.from_numpy(cls)
# [batch_size, channel, width, height]
# [batch_size, all_anchors, num_classes+1]
# [batch_size, all_anchors, num_values+1]
return image_batch, cls, reg
def compute_targets(self, image_group, annotations_group):
"""
Compute target outputs for the network using images and their annotations.
"""
"""
Compute target outputs for the network using images and their annotations.
"""
batches_targets = anchor_targets_bbox(self.anchors, image_group,
annotations_group,
self.num_classes())
return list(batches_targets)
def compute_targets(self, image_group, annotations_group):
"""
Compute target outputs for the network using images and their annotations.
Args:
image_group: List with images of a group/batch
annotations_group: List with annotations of a group/batch
Returns:
List with the target batches for EfficientPose
"""
batches_targets = anchor_targets_bbox(
self.anchors,
image_group,
annotations_group,
num_classes=self.num_classes(),
num_rotation_parameters=self.rotation_parameter +
2, #+1 for the is_symmetric flag and +1 for the class idx to choose the correct model 3d points
num_translation_parameters=self.
translation_parameter, #x,y in 2D and Tz
translation_anchors=self.translation_anchors,
)
return list(batches_targets)
