def read_COCOA(ann, h, w):
if 'visible_mask' in ann.keys():
rle = [ann['visible_mask']]
else:
rles = maskUtils.frPyObjects([ann['segmentation']], h, w)
rle = maskUtils.merge(rles)
modal = maskUtils.decode(rle).squeeze()
if np.all(modal != 1):
# if the object if fully occluded by others,
# use amodal bbox as an approximated location,
# note that it will produce random amodal results.
amodal = maskUtils.decode(maskUtils.merge(
maskUtils.frPyObjects([ann['segmentation']], h, w)))
bbox = utils.mask_to_bbox(amodal)
else:
bbox = utils.mask_to_bbox(modal)
return modal, bbox, 1 # category as constant 1
def get_instance(self, idx, with_gt=False):
assert not with_gt, \
"Mapillary Vista has no ground truth for ordering or amodal masks."
imgidx, regidx = self.indexing[idx]
# img
image_id = self.annot_info[imgidx]['image_id']
image_fn = image_id + ".jpg"
# region
instance_map = np.array(
Image.open("{}/instances/{}.png".format(
self.root, image_id)), dtype=np.uint16)
h, w = instance_map.shape[:2]
reg_info = self.annot_info[imgidx]['regions'][regidx]
modal = (instance_map == reg_info['instance_id']).astype(np.uint8)
category = reg_info['category_id']
bbox = np.array(utils.mask_to_bbox(modal))
return modal, bbox, category, image_fn, None
def make_KINS_output(self, idx, amodal_pred, category):
results = []
for i in range(amodal_pred.shape[0]):
data = dict()
rle = maskUtils.encode(
np.array(amodal_pred[i, :, :, np.newaxis], order='F'))[0]
if hasattr(self.data_reader, 'img_ids'):
data['image_id'] = self.data_reader.img_ids[idx]
data['category_id'] = category[i].item()
if isinstance(rle['counts'], bytes):
rle['counts'] = rle['counts'].decode()
data['segmentation'] = rle
data['bbox'] = utils.mask_to_bbox(amodal_pred[i, :, :])
data['area'] = float(data['bbox'][2] * data['bbox'][3])
data['iscrowd'] = 0
data['score'] = 1.
data['id'] = self.count
results.append(data)
self.count += 1
return results
def get_image_instances(self, idx, with_gt=False, with_anns=False, ignore_stuff=False):
assert not with_gt
assert not ignore_stuff
# img
image_id = self.annot_info[idx]['image_id']
image_fn = image_id + ".jpg"
# region
instance_map = np.array(
Image.open("{}/instances/{}.png".format(
self.root, image_id)), dtype=np.uint16)
h, w = instance_map.shape[:2]
instance_ids = np.unique(instance_map)
category = instance_ids // 256
num_instance = len(instance_ids)
instance_ids_tensor = np.zeros((num_instance, h, w), dtype=np.uint16)
instance_ids_tensor[...] = instance_ids[:, np.newaxis, np.newaxis]
modal = (instance_ids_tensor == instance_map).astype(np.uint8)
bboxes = []
for i in range(modal.shape[0]):
bboxes.append(utils.mask_to_bbox(modal[i,...]))
return modal, category, np.array(bboxes), None, image_fn
def objectSave(self):
obj = self.objects[self.this_obj - 1]
crop_obj = utils.crop_padding(obj,
utils.mask_to_bbox(obj[:, :, 3]),
pad_value=(0, 0, 0, 0))
self.window().objectSaveAs(crop_obj)