def init_target_boxes(self):
"""Get the target bounding boxes for the initial augmented samples."""
self.classifier_target_box = self.get_iounet_box(
self.pos, self.target_sz, self.init_sample_pos,
self.init_sample_scale)
init_target_boxes = TensorList()
for T in self.transforms:
init_target_boxes.append(
self.classifier_target_box +
torch.Tensor([T.shift[1], T.shift[0], 0, 0]))
init_target_boxes = torch.cat(init_target_boxes.view(1, 4),
0).to(self.p.device)
self.target_boxes = init_target_boxes.new_zeros(
self.p.sample_memory_size, 4)
self.target_boxes[:init_target_boxes.shape[0], :] = init_target_boxes
return init_target_boxes
def init_iou_net(self, backbone_feat):
# Setup IoU net and objective
for p in self.net.bb_regressor.parameters():
p.requires_grad = False
# Get target boxes for the different augmentations
self.classifier_target_box = self.get_iounet_box(
self.pos, self.target_sz, self.init_sample_pos,
self.init_sample_scale)
target_boxes = TensorList()
if self.p.iounet_augmentation:
for T in self.transforms:
if not isinstance(
T, (augmentation.Identity, augmentation.Translation,
augmentation.FlipHorizontal,
augmentation.FlipVertical, augmentation.Blur)):
break
target_boxes.append(
self.classifier_target_box +
torch.Tensor([T.shift[1], T.shift[0], 0, 0]))
else:
target_boxes.append(self.classifier_target_box + torch.Tensor([
self.transforms[0].shift[1], self.transforms[0].shift[0], 0, 0
]))
target_boxes = torch.cat(target_boxes.view(1, 4), 0).to(self.p.device)
# Get iou features
iou_backbone_feat = self.get_iou_backbone_features(backbone_feat)
# Remove other augmentations such as rotation
iou_backbone_feat = TensorList(
[x[:target_boxes.shape[0], ...] for x in iou_backbone_feat])
# Get modulation vector
self.iou_modulation = self.get_iou_modulation(iou_backbone_feat,
target_boxes)
self.iou_modulation = TensorList(
[x.detach().mean(0) for x in self.iou_modulation])