def merge_aug_results(self, aug_bboxes, aug_poses, aug_scores, aug_areas,
aug_vis, img_metas):
recovered_bboxes = []
recovered_poses = []
recovered_areas = []
for bboxes, poses, areas, img_info in zip(aug_bboxes, aug_poses,
aug_areas, img_metas):
img_shape = img_info[0]['img_shape']
scale_factor = img_info[0]['scale_factor']
flip = img_info[0]['flip']
# assert not flip
bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip)
if flip:
kpt_name, kpt_flip_map = get_keypoints()
poses = poses.view(poses.shape[0], -1, 2)
poses = flip_keypoints(kpt_name, kpt_flip_map, poses,
img_shape[1])
poses = poses.view(poses.shape[0], -1)
poses = poses / scale_factor
areas = areas / (scale_factor * scale_factor)
recovered_bboxes.append(bboxes)
recovered_poses.append(poses)
recovered_areas.append(areas)
bboxes = torch.cat(recovered_bboxes, dim=0)
poses = torch.cat(recovered_poses, dim=0)
areas = torch.cat(recovered_areas, dim=0)
scores = torch.cat(aug_scores, dim=0)
vis = torch.cat(aug_vis, dim=0)
return bboxes, poses, areas, scores, vis
def merge_aug_results(self, aug_bboxes, aug_scores, img_metas):
"""Merge augmented detection bboxes and scores.
Args:
aug_bboxes (list[Tensor]): shape (n, 4*#class)
aug_scores (list[Tensor] or None): shape (n, #class)
img_shapes (list[Tensor]): shape (3, ).
Returns:
tuple: (bboxes, scores)
"""
recovered_bboxes = []
for bboxes, img_info in zip(aug_bboxes, img_metas):
img_shape = img_info[0]['img_shape']
scale_factor = img_info[0]['scale_factor']
flip = img_info[0]['flip']
flip_direction = img_info[0]['flip_direction']
bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip,
flip_direction)
recovered_bboxes.append(bboxes)
bboxes = torch.cat(recovered_bboxes, dim=0)
if aug_scores is None:
return bboxes
else:
scores = torch.cat(aug_scores, dim=0)
return bboxes, scores
def merge_aug_results(self, aug_results, img_metas):
"""Merge augmented detection bboxes and score.
Args:
aug_results (list[list[Tensor]]): Det_bboxes and det_labels of each
image.
img_metas (list[list[dict]]): Meta information of each image, e.g.,
image size, scaling factor, etc.
Returns:
tuple: (bboxes, labels)
"""
recovered_bboxes, aug_labels = [], []
for bboxes_labels, img_info in zip(aug_results, img_metas):
img_shape = img_info[0]['img_shape'] # using shape before padding
scale_factor = img_info[0]['scale_factor']
flip = img_info[0]['flip']
bboxes, labels = bboxes_labels
bboxes, scores = bboxes[:, :4], bboxes[:, -1:]
bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip)
recovered_bboxes.append(torch.cat([bboxes, scores], dim=-1))
aug_labels.append(labels)
bboxes = torch.cat(recovered_bboxes, dim=0)
labels = torch.cat(aug_labels)
if bboxes.shape[0] > 0:
out_bboxes, out_labels = self.bbox_head._bboxes_nms(
bboxes, labels, self.bbox_head.test_cfg)
else:
out_bboxes, out_labels = bboxes, labels
return out_bboxes, out_labels
def merge_aug_vote_results(self, aug_bboxes, aug_labels, img_metas):
"""Merge augmented detection bboxes and scores.
Args:
aug_bboxes (list[Tensor]): shape (n, 4*#class)
aug_scores (list[Tensor] or None): shape (n, #class)
img_shapes (list[Tensor]): shape (3, ).
rcnn_test_cfg (dict): rcnn test config.
Returns:
tuple: (bboxes, scores)
"""
recovered_bboxes = []
for bboxes, img_info in zip(aug_bboxes, img_metas):
img_shape = img_info[0]['img_shape']
scale_factor = img_info[0]['scale_factor']
flip = img_info[0]['flip']
bboxes[:, :4] = bbox_mapping_back(bboxes[:, :4], img_shape, scale_factor, flip)
recovered_bboxes.append(bboxes)
bboxes = torch.cat(recovered_bboxes, dim=0)
if aug_labels is None:
return bboxes
else:
labels = torch.cat(aug_labels, dim=0)
return bboxes, labels
def merge_aug_results(self, aug_bboxes, aug_masks, aug_scores, img_metas):
recovered_bboxes = []
recovered_masks = []
for bboxes, masks, img_info in zip(aug_bboxes, aug_masks, img_metas):
img_shape = img_info[0]['img_shape']
scale_factor = img_info[0]['scale_factor']
flip = img_info[0]['flip']
bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip)
masks = bbox_mapping_back(masks, img_shape, scale_factor, flip)
recovered_bboxes.append(bboxes)
recovered_masks.append(masks)
bboxes = torch.cat(recovered_bboxes, dim=0)
masks = torch.cat(recovered_masks, dim=0)
if aug_scores is None:
return bboxes, masks
else:
scores = torch.cat(aug_scores, dim=0)
return bboxes, masks, scores
def merge_aug_results(self, aug_bboxes, aug_scores, aug_points, aug_masks, img_metas):
"""
Merge augmented detection bboxes and scores.
Args:
aug_bboxes (list[Tensor]): shape (n, 4*#class)
aug_scores (list[Tensor] or None): shape (n, #class)
aug_points (list[Tensor]): shape (n, pts_num*#class)
img_shapes (list[Tensor]): shape (3, ).
rcnn_test_cfg (dict): rcnn test config.
Returns:
tuple: (bboxes, scores)
"""
def points_mapping_back(points, img_shape, scale_factor, flip):
flipped = points.clone()
if flip:
flipped[:, 0] = img_shape[1] - points[:, 0]
flipped = flipped / scale_factor
return flipped
recovered_bboxes = []
recovered_points = []
for bboxes, points, img_info in zip(aug_bboxes, aug_points, img_metas):
img_shape = img_info[0]['img_shape']
scale_factor = img_info[0]['scale_factor']
flip = img_info[0]['flip']
bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip)
points = points_mapping_back(points, img_shape, scale_factor, flip)
recovered_bboxes.append(bboxes)
recovered_points.append(points)
bboxes = torch.cat(recovered_bboxes, dim=0)
points = torch.cat(recovered_points, dim=0)
if aug_scores is None:
return bboxes
else:
scores = torch.cat(aug_scores, dim=0)
masks = torch.cat(aug_masks, dim=0)
return bboxes, scores, points, masks
