Potentially Faster CCL Extraction

[william-silversmith]

Hi!

I'm the author of cc3d and from time to time I search GitHub to see how people are using the library. I noticed that there are a few places that could potentially see some improvement even though you clearly made use of many of cc3d's tricks.

EfficientSegmentation/Common/mask_process.py

Lines 85 to 112 in 5a8eda5

	def remove_small_connected_object(npy_mask, area_least, out_mask, out_label):
	labels_out = cc3d.connected_components(npy_mask, connectivity=26)
	for label, extracted in cc3d.each(labels_out, binary=True, in_place=True):
	area = np.sum(extracted)
	if area > area_least:
	out_mask[labels_out == int(label)] = out_label
	def extract_topk_largest_candidates(npy_mask: np.array, out_num_label: List, area_least: int) -> np.array:
	mask_shape = npy_mask.shape
	out_mask = np.zeros([mask_shape[1], mask_shape[2], mask_shape[3]], np.uint8)
	for i in range(mask_shape[0]):
	t_mask = npy_mask[i].copy()
	keep_topk_largest_connected_object(t_mask, out_num_label[i], area_least, out_mask, i+1)
	return out_mask
	def keep_topk_largest_connected_object(npy_mask, k, area_least, out_mask, out_label):
	labels_out = cc3d.connected_components(npy_mask, connectivity=26)
	areas = {}
	for label, extracted in cc3d.each(labels_out, binary=True, in_place=True):
	areas[label] = fastremap.foreground(extracted)
	candidates = sorted(areas.items(), key=lambda item: item[1], reverse=True)
	for i in range(min(k, len(candidates))):
	if candidates[i][1] > area_least:
	out_mask[labels_out == int(candidates[i][0])] = out_label

cc3d has some very efficient functions for performing these two tasks. cc3d.dust and cc3d.largest_k. You might be surprised at the improvement.

Anywho, if this isn't helpful to your project feel free to ignore these tips. Thanks for using cc3d!

Will

[Shanghai-Aitrox-Technology]

@william-silversmith 3Q for your comments.