Unlabeled Images Mask in Split

[liheyoung]

Unlabeled Images Mask in Split

[LiheYoung]

@tanveer6715 Add a condition:

if self.mode == 'train_u':
    mask = Image.fromarray(np.zeros((img.size[1], img.size[0]), dtype=np.uint8))
else:
    mask = Image.fromarray(np.array(Image.open(os.path.join(self.root, id.split(' ')[1])))) 

[LiheYoung]

We do not use manual annotations of unlabeled images for training. The masks of unlabeled images are used to record padded regions during image pre-processing.

If you do not have masks, you can simply duplicate the image paths in each line of split file, for example:
path/of/image1.jpg path/of/image1.jpg

Then you need to modify this line:

UniMatch/dataset/semi.py

Line 35 in a5d2a52

mask = Image.fromarray(np.array(Image.open(os.path.join(self.root, id.split(' ')[1]))))

to:

mask = Image.fromarray(np.zeros((img.size[1], img.size[0]), dtype=np.uint8))

[rahulsnkr]

thanks for the clarification!

[tanveer6715]

We do not use manual annotations of unlabeled images for training. The masks of unlabeled images are used to record padded regions during image pre-processing.

If you do not have masks, you can simply duplicate the image paths in each line of split file, for example: path/of/image1.jpg path/of/image1.jpg

Then you need to modify this line:

UniMatch/dataset/semi.py

Line 35 in a5d2a52

mask = Image.fromarray(np.array(Image.open(os.path.join(self.root, id.split(' ')[1]))))

to:

mask = Image.fromarray(np.zeros((img.size[1], img.size[0]), dtype=np.uint8))

Hi,
This process the data in the training when we do not want to use labels in the unlabeled data split but in the validation step after each epoch the model shows 0 miou as it do not use masks in the validation step as well to compare the predictions. How to tackle this problem?