[feat] Adding camvid segmentation multiclass as an example

[TimbusCalin]

• Multiclass semantic segmentation example on all classes
• Background class is added as an additional class
• Initial PyTorch Lightning example is maintained, adapted in accordance with multiclass

Fixes: #911

[qubvel]

Hi @TimbusCalin, thanks a lot for your contribution! Looks great overall and it's pretty close to being merged, here are just a few minor comments:

1. Remove commented code or replace it with comments / assertations

# Specify classes only if necessary
# CLASSES = ['road','car']

# Debugging unique values and type of mask
# print(f"Mask dtype after casting: {mask.dtype}, unique values: {torch.unique(mask)}, shape: {mask.shape}")

# Debugging logits shape
# print(f"Logits mask shape: {logits_mask.shape}")

maybe somewhere else

2. It would also be cool to add metric computation per class, if you have time and know how to do that :) but that's optional

[abdksyed]

I think the OUT_CLASSES should be equal to length of CLASSES. Since the background(unlabelled) classes is already included in the CLASSES list.

OUT_CLASSES = len(train_dataset.CLASSES) + 1

This should be in case when no class is specified.

OUT_CLASSES = len(train_dataset.CLASSES)

The +1 will happen when the user specifies a list of classes, at that point we will have to use the provided len(class_list) + 1.

[TimbusCalin]

Hello @abdksyed @qubvel.

I am fine with this approach that @abdksyed mentioned, to integrate my PR now, and I will let @abdksyed do the class-metrics-wise function now, once the PR is integrated; glad that we synchronized on this.

@abdksyed You are right, that's why I updated it accordingly in this commit here aa8226c

[abdksyed]

Hello @abdksyed @qubvel.

I am fine with this approach that @abdksyed mentioned, to integrate my PR now, and I will let @abdksyed do the class-metrics-wise function now, once the PR is integrated; glad that we synchronized on this.

@abdksyed You are right, that's why I updated it accordingly in this commit here aa8226c

Oh sorry, I missed your last commit. Apologies

[TimbusCalin]

@abdksyed no issue, another pair of (good!) eyes is always necessary and welcome!

[TimbusCalin]

@qubvel I guess the dillema is solved. We can integrate if you have no other remarks.

[qubvel]

Thanks for your contribution!

[sajeedmehrab]

@abdksyed Do you have an example for how we can compute the class-wise miou?

[abdksyed]

@sajeedmehrab
I use the torchmetrics. You can give the per_class as True to get class wise iou.

[qubvel]

Would appreciate if you contribute torch metrics example 😊