def compute_BD(A, B):
bd = 0.
for i in range(len(A)):
gt_mask = maskUtils.decode(A[i]["segmentation"])
max_dice = 0.
for j in range(len(B)):
pred_mask = maskUtils.decode(B[j]["segmentation"])
dice = compute_dice(pred_mask, gt_mask)
max_dice = max(max_dice, dice)
bd += max_dice
return bd / len(A)
def annList2mask(annList):
n_anns = len(annList)
if n_anns == 0:
return {"mask": None}
ann = annList[0]
try:
h, w = ann["mask"].shape
except:
h, w = ann["height"], ann["width"]
mask = np.zeros((h, w), int)
for i in range(n_anns):
ann = annList[i]
if "mask" in ann:
ann_mask = ann["mask"]
else:
ann_mask = maskUtils.decode(ann["segmentation"])
assert ann_mask.max() <= 1
mask += ann_mask
# mask[mask==1] = ann["category_id"]
return {"mask": mask}
def annList2mask(annList, box=False, color=False):
n_anns = len(annList)
if n_anns == 0:
return {"mask":None}
ann = annList[0]
try:
h, w = ann["mask"].shape
except:
h, w = ann["height"], ann["width"]
mask = np.zeros((h, w), int)
for i in range(n_anns):
ann = annList[i]
if "mask" in ann:
ann_mask = ann["mask"]
else:
ann_mask = maskUtils.decode(ann["segmentation"])
assert ann_mask.max() <= 1
if color:
mask[ann_mask!=0] = i + 1
else:
mask += ann_mask
for i in range(n_anns):
if box:
box_mask = ann2bbox(annList[i])["mask"]
mask[box_mask!=0] = 255
# mask[mask==1] = ann["category_id"]
return {"mask":mask}
def annList2maskList(annList, box=False, color=False):
n_anns = len(annList)
if n_anns == 0:
return {"mask":None}
ann = annList[0]
try:
h, w = ann["mask"].shape
except:
h, w = ann["height"], ann["width"]
maskList = np.zeros((h, w, n_anns), int)
categoryList = np.zeros(n_anns, int)
for i in range(n_anns):
ann = annList[i]
if "mask" in ann:
ann_mask = ann["mask"]
else:
ann_mask = maskUtils.decode(ann["segmentation"])
assert ann_mask.max() <= 1
maskList[:,:,i] = ann_mask
categoryList[i] = ann["category_id"]
# mask[mask==1] = ann["category_id"]
return {"maskList":maskList, "categoryList":categoryList}
def compute_coverage(A, B):
mucov = 0.
mwcov = 0.
gt_size = 0.
for i in range( len(A)):
gt_mask = maskUtils.decode(A[i]["segmentation"])
best_iou = 0.
for j in range(len(B)):
pred_mask = maskUtils.decode(B[j]["segmentation"])
iou=compute_iou(pred_mask, gt_mask)
best_iou = max(best_iou, iou)
mucov += best_iou
mwcov += best_iou * gt_mask.sum()
gt_size += gt_mask.sum()
return (mucov / len(A)), (mwcov / gt_size), gt_size
def __getitem__(self, i):
encoded = self.proposals[i]["segmentation"]
proposal_mask = maskUtils.decode(encoded)
return {"mask": proposal_mask, "score": self.proposals[i]["score"]}
def ann2mask(ann):
mask = maskUtils.decode(ann["segmentation"])
# mask[mask==1] = ann["category_id"]
return {"mask":mask}
def __getitem__(self, index):
name_id = self.ids[index]
name = self.image_names[index]
image = np.array(
Image.open(self.path +
"/{}/{}".format(self.split +
self.year, name)).convert('RGB'))
points = np.zeros(image.shape[:2], "uint8")[:, :, None]
counts = np.zeros(80)
maskVoid = np.zeros(points.shape[:2])
annList = ms.load_pkl(
self.path + "/groundtruth/{}_{}.pkl".format(self.split, name))
h, w, _ = image.shape
maskClasses = np.zeros((h, w), int)
maskObjects = np.zeros((h, w), int)
for obj_id, ann in enumerate(annList):
mask = maskUtils.decode(COCO.annToRLE_issam(h, w, ann))
if ann["iscrowd"]:
maskVoid += mask
else:
dist = distance_transform_edt(mask)
yx = np.unravel_index(dist.argmax(), dist.shape)
label = self.category2label[int(ann["category_id"])]
points[yx] = label
counts[label - 1] += 1
assert mask.max() <= 1
mask_ind = mask == 1
maskObjects[mask_ind] = obj_id + 1
maskClasses[mask_ind] = label
maskVoid = maskVoid.clip(0, 1)
assert maskVoid.max() <= 1
counts = torch.LongTensor(counts)
image, points, maskObjects, maskClasses = self.transform_function(
[image, points, maskObjects, maskClasses])
# Sharp Proposals
image_id = int(name[:-4].split("_")[-1])
SharpProposals_name = self.proposals_path + "{}".format(image_id)
lcfcn_pointList = self.get_lcfcn_pointList(str(image_id))
assert image_id == name_id
if self.split == "train":
return {
"images": image,
"points": points.squeeze(),
"SharpProposals_name": str(name_id),
"counts": counts,
"index": index,
"name": str(name_id),
"image_id": str(image_id),
"maskObjects": maskObjects * 0,
"maskClasses": maskClasses * 0,
"proposals_path": self.proposals_path,
"dataset": "coco2014",
"lcfcn_pointList": lcfcn_pointList,
"split": self.split
}
else:
return {
"images": image,
"points": points.squeeze(),
"SharpProposals_name": str(name_id),
"counts": counts,
"index": index,
"name": str(name_id),
"image_id": str(image_id),
"maskObjects": maskObjects,
"maskClasses": maskClasses,
"proposals_path": self.proposals_path,
"dataset": "coco2014",
"lcfcn_pointList": lcfcn_pointList,
"split": self.split
}