Our code snapshot:
cur_image = cv2.imread(cur_image_path, -1)
cur_label = cv2.imread(cur_label_path, -1)
if cur_image.shape[0] != cur_label.shape[0] or cur_image.shape[1] != cur_label.shape[1]:
print('Warning Image shape:{} not equal to Label shape:{}'.format(cur_image.shape, cur_label.shape))
continue
Running Results:
FloodNet/Train/Labeled/Flooded/image/7556.jpg FloodNet/Train/Labeled/Flooded/mask/7556_lab.png
Warning Image shape:(3072, 4592, 3) not equal to Label shape:(3000, 4000)
FloodNet/Train/Labeled/Flooded/image/7485.jpg FloodNet/Train/Labeled/Flooded/mask/7485_lab.png
Warning Image shape:(3072, 4592, 3) not equal to Label shape:(3000, 4000)
FloodNet/Train/Labeled/Flooded/image/7524.jpg FloodNet/Train/Labeled/Flooded/mask/7524_lab.png
Warning Image shape:(3072, 4592, 3) not equal to Label shape:(3000, 4000)
FloodNet/Train/Labeled/Flooded/image/7298.jpg FloodNet/Train/Labeled/Flooded/mask/7298_lab.png
Warning Image shape:(3072, 4592, 3) not equal to Label shape:(3000, 4000)
FloodNet/Train/Labeled/Flooded/image/7330.jpg FloodNet/Train/Labeled/Flooded/mask/7330_lab.png
Warning Image shape:(3072, 4592, 3) not equal to Label shape:(3000, 4000)
FloodNet/Train/Labeled/Flooded/image/9003.jpg FloodNet/Train/Labeled/Flooded/mask/9003_lab.png
FloodNet/Train/Labeled/Flooded/image/7321.jpg FloodNet/Train/Labeled/Flooded/mask/7321_lab.png
Warning Image shape:(3072, 4592, 3) not equal to Label shape:(3000, 4000)
FloodNet/Train/Labeled/Flooded/image/6716.jpg FloodNet/Train/Labeled/Flooded/mask/6716_lab.png
Warning Image shape:(3072, 4592, 3) not equal to Label shape:(3000, 4000)
FloodNet/Train/Labeled/Flooded/image/7267.jpg FloodNet/Train/Labeled/Flooded/mask/7267_lab.png
Warning Image shape:(3072, 4592, 3) not equal to Label shape:(3000, 4000)
FloodNet/Train/Labeled/Flooded/image/7463.jpg FloodNet/Train/Labeled/Flooded/mask/7463_lab.png
Warning Image shape:(3072, 4592, 3) not equal to Label shape:(3000, 4000)
Some of the raw images are not necessarily 3000x4000, but all the original masks are 3000x4000. So you have to resize the predicted mask to 3000x4000 to compare with the original masks.
If the participants submit predicted masks with size different than 3000x4000, our evaluation script will resize them to 3000x4000 and then will calculate the mIoU. Hope this answers your question.
Thank you!
Posted by: binalab @ April 6, 2021, 3:46 p.m.