def convertImages_fast(predictionList, groundTruthList, args, mapping_dict):
df = pd.DataFrame(columns=[
'ImageId', 'LabelId', 'Confidence', 'PixelCount', 'EncodedPixels'
])
df_count = 0
for list_index, filename in enumerate(groundTruthList):
imageID = mapping_dict[filename]
if predictionList[list_index]:
predicitionFile = open(predictionList[list_index], "r")
predictionlines = predicitionFile.readlines()
for predictionline in predictionlines:
predictionInfo = predictionline.split(' ')
img = Image.open(predictionInfo[0])
InstanceMap = np.array(img)
InstanceMap = InstanceMap / np.max(InstanceMap)
# Historical code
LabelId = int(predictionInfo[1])
Confidence = float(predictionInfo[2].split('\n')[0])
idmap1d = np.reshape(InstanceMap > 0, (-1))
PixelCount = np.sum(idmap1d)
EncodedPixels = rle_encoding(InstanceMap)
df.loc[df_count] = [
imageID, LabelId, Confidence, PixelCount, EncodedPixels
]
df_count += 1
df.to_csv(args.csv_file, header=True, index=False)
print('Finish converting file: %s' % args.csv_file)
return
def convertImages_with_image(filename, list_index, predictionList, args,
mapping_dict):
df = pd.DataFrame(columns=[
'ImageId', 'LabelId', 'Confidence', 'PixelCount', 'EncodedPixels'
])
df_count = 0
imageID = mapping_dict[filename]
if os.path.exists(predictionList[list_index]):
predicitionFile = open(predictionList[list_index], "r")
predictionlines = predicitionFile.readlines()
for predictionline in predictionlines:
predictionInfo = predictionline.split(' ')
LabelId = int(predictionInfo[1])
Confidence = float(predictionInfo[2].split('\n')[0])
img = Image.open(predictionInfo[0])
InstanceMap = np.array(img)
idmap1d = np.reshape(InstanceMap > 0, (-1))
PixelCount = np.sum(idmap1d)
EncodedPixels = rle_encoding(idmap1d)
df.loc[df_count] = [
imageID, LabelId, Confidence, PixelCount, EncodedPixels
]
df_count += 1
df.to_csv(args.csv_file_image, header=True, index=False)
print('Finish converting file: %s.' % (args.csv_file_image))
def convertImages_fast_old(predictionList, groundTruthList, args,
mapping_dict):
csv = open(args.csv_file, 'a')
dataset = WAD_CVPR2018('/media/samsumg_1tb/CVPR2018_WAD')
for list_index, filename in enumerate(groundTruthList):
imageID = mapping_dict[filename]
if predictionList[list_index]:
predicitionFile = open(predictionList[list_index], "r")
predictionlines = predicitionFile.readlines()
for predictionline in predictionlines:
predictionInfo = predictionline.split(' ')
img = Image.open(predictionInfo[0])
InstanceMap = np.array(img)
csv.write("{},".format(imageID))
# Historical code
if int(predictionInfo[1]
) in dataset.contiguous_category_id_to_json_id.keys():
csv.write("{},".format(
dataset.contiguous_category_id_to_json_id[int(
predictionInfo[1])]))
else:
csv.write("{},".format(predictionInfo[1]))
csv.write("{},".format(predictionInfo[2].split('\n')[0]))
idmap1d = np.reshape(InstanceMap == 1, (-1))
Totalcount = np.sum(idmap1d)
csv.write("{},".format(Totalcount))
EncodedPixed = rle_encoding(InstanceMap)
csv.write(EncodedPixed)
csv.write('\n')
csv.close()
print('Finish converting file: %s' % args.csv_file)
return
def convertImages_fast(predictionList, groundTruthList, args, mapping_dict):
csv = open(args.csv_file, 'a')
for list_index, filename in enumerate(groundTruthList):
imageID = mapping_dict[filename]
predicitionFile = open(predictionList[list_index], "r")
predictionlines = predicitionFile.readlines()
for predictionline in predictionlines:
predictionInfo = predictionline.split(' ')
img = Image.open(predictionInfo[0])
InstanceMap = np.array(img)
csv.write("{},".format(imageID))
csv.write("{},".format(predictionInfo[1]))
csv.write("{},".format(predictionInfo[2].split('\n')[0]))
idmap1d = np.reshape(InstanceMap == 1, (-1))
Totalcount = np.sum(idmap1d)
csv.write("{},".format(Totalcount))
EncodedPixed = rle_encoding(InstanceMap)
csv.write(EncodedPixed)
csv.close()
return
def convertImages_with_postprocessing_image(filename, list_index,
predictionList, args,
mapping_dict):
df = pd.DataFrame(columns=[
'ImageId', 'LabelId', 'Confidence', 'PixelCount', 'EncodedPixels'
])
df_count = 0
del_count = 0
imageID = mapping_dict[filename]
if os.path.exists(predictionList[list_index]):
predicitionFile = open(predictionList[list_index], "r")
predictionlines = predicitionFile.readlines()
# We keep a mask for the whole image, and fill it with masks. The maskes are filled
# with descending order and if there is an overlap, we discard such instance.
img_mask_list = []
img_mask_valid = []
conf_list = []
label_list = []
for predictionline in predictionlines:
predictionInfo = predictionline.split(' ')
img_mask_list.append(predictionInfo[0])
label_list.append(int(predictionInfo[1]))
conf_list.append(float(predictionInfo[2].split('\n')[0]))
conf_order = np.argsort(-np.array(conf_list))
image_mask_all = np.zeros(shape=(2710, 3384))
for conf_idx in conf_order:
# skip flag indicates whether we will skip this instance
skip_flag = False
img = Image.open(img_mask_list[conf_idx])
InstanceMap = np.array(img)
image_mask_all += InstanceMap > 0
if len(np.unique(image_mask_all)) > 2:
# we will check whether we will keep this instance by looping over all previous instances:
for im_mask in img_mask_valid:
rle1 = maskUtils.encode(
np.array(im_mask[:, :, np.newaxis], order='F'))[0]
rle2 = maskUtils.encode(
np.array(InstanceMap[:, :, np.newaxis], order='F'))[0]
iou = maskUtils.iou([rle1], [rle2], [0])
if iou[0][0] > args.del_overlap:
image_mask_all -= InstanceMap
del_count += 1
skip_flag = True
continue
if not skip_flag:
img_mask_valid.append(InstanceMap)
LabelId = label_list[conf_idx]
Confidence = conf_list[conf_idx]
idmap1d = np.reshape(InstanceMap > 0, (-1))
PixelCount = np.sum(idmap1d)
EncodedPixels = rle_encoding(InstanceMap)
df.loc[df_count] = [
imageID, LabelId, Confidence, PixelCount, EncodedPixels
]
df_count += 1
df.to_csv(args.csv_file_image, header=True, index=False)
print('Finish converting file: %s with %d deleting overlaps' %
(args.csv_file_image, del_count))
