def overwrite_label_img_masks(img_fpath: str, label_img_fpath: str,
parent_fname_to_updatelist_dict, mld: Any,
classname_to_id_map: Mapping[str, int],
require_strict_boundaries: bool,
split: str) -> None:
"""
Swap the pixel values inside a label map's mask to a new value. This
effectively changes the mask's category on disk.
Get fname stem from rgb image file path
Get sequence ID/parent from label image path file system parent.
Args:
- img_fpath:
- label_img_fpath:
- parent_fname_to_updatelist_dict:
- mld:
- classname_to_id_map:
- require_strict_boundaries:
- split:
Returns:
- None
"""
fname_stem = Path(img_fpath).stem
parent = Path(label_img_fpath).parts[
-2] # get parent name, functions as sequence ID
if fname_stem not in parent_fname_to_updatelist_dict[parent]:
# we loop through the entire dataset, and many images won't have any
# updated masks. they'll pass through here.
return
# load up each label_img
# label image will already by in universal taxonomy
label_img = imageio.imread(label_img_fpath)
update_records = parent_fname_to_updatelist_dict[parent][fname_stem]
for rec in update_records:
# if it is, perform each update as described in the object
segment_mask = mld.get_segment_mask(parent, rec.segmentid, fname_stem,
split)
if segment_mask is None:
print('No such mask found, exiting...')
exit()
# update the label image each time
orig_class_idx = classname_to_id_map[rec.orig_class]
new_class_idx = classname_to_id_map[rec.relabeled_class]
label_img = swap_px_inside_mask(label_img, segment_mask,
orig_class_idx, new_class_idx,
require_strict_boundaries)
# save it to disk, at new data root, using same abs path as before
overwrite = True # False #
if overwrite:
imageio.imwrite(label_img_fpath, label_img)
def test_swap_px_inside_mask():
""" """
segment_mask = np.array([[0, 0, 1, 0], [0, 1, 1, 0], [1, 1, 1, 1]])
semantic_img = np.zeros((3, 4), dtype=np.uint8)
new_img = swap_px_inside_mask(semantic_img,
segment_mask,
old_val=0,
new_val=8,
require_strict_boundaries=True)
gt_new_img = np.array([[0, 0, 8, 0], [0, 8, 8, 0], [8, 8, 8, 8]])
assert np.allclose(gt_new_img, new_img)
def test_swap_px_inside_mask():
label_img = np.array([[7, 8, 9, 0], [0, 7, 0, 7], [3, 4, 7, 1]],
dtype=np.uint8)
segment_mask = np.array([[1, 0, 0, 0], [0, 1, 0, 1], [0, 0, 1, 0]],
dtype=np.uint8)
old_val = 7
new_val = 255
new_label_img = swap_px_inside_mask(label_img, segment_mask, old_val,
new_val)
gt_new_label_img = np.array(
[[255, 8, 9, 0], [0, 255, 0, 255], [3, 4, 255, 1]], dtype=np.uint8)
assert np.allclose(new_label_img, gt_new_label_img)
def write_semantic_from_panoptic(
cse: COCOSemanticExtractor,
split: str,
instance_img_fpath: str,
ignore_idx: int = 255
) -> None:
"""
Args:
- cse
- split
- instance_img_fpath
- ignore_idx
Returns:
- None
"""
fname_stem = Path(instance_img_fpath).stem
instance_id_img = cse.instance_api.get_instance_id_img(split, fname_stem)
img_annot = cse.semantic_api.get_img_annotation(split, fname_stem)
# default pixel value is unlabeled
semantic_img = np.ones_like(instance_id_img, dtype=np.uint8) * ignore_idx
for segment in img_annot['segments_info']:
segmentid = segment['id']
categoryid = segment['category_id']
segment_mask = (instance_id_img == segmentid).astype(np.uint8)
semantic_img = swap_px_inside_mask(
semantic_img,
segment_mask,
old_val=ignore_idx,
new_val=categoryid,
require_strict_boundaries=True
)
semantic_fpath = instance_img_fpath.replace(
f'annotations/panoptic_{split}2017',
f'semantic_annotations201/{split}2017' # in 201-class taxonomy
)
check_mkdir(Path(semantic_fpath).parent)
imageio.imwrite(semantic_fpath, semantic_img)