ChangeDetectionTask - how to implement

[robmarkcole]

As noted in issue #336 there is not currently a ChangeDetectionTask I am seeking guidance on how to implement this task, or instruction on how SegmentationTask or similar can be utilised for change detection.
Thanks in advance

[adamjstewart]

SemanticSegmentationTask could be utilized by stacking all images along the channel dimension: B x T x C x H x W -> B x (T x C) x H x W

I also wouldn't be opposed to having a dedicated ChangeDetectionTask that takes samples with "image1" and "image2" keys. We mostly just didn't settle on a naming scheme. I think it also depends on whether or not there are custom models that only make sense for change detection and require images to be separate.

[robmarkcole]

One model I am interested in is TinyCD which takes pairs of images - there also might be situations where you would want to apply a transform on one of the pair but not the other (perhaps a seasonal augmentation).

Looking at OSCD the images are stacked, so would need to be unstacked for tinycd

[adamjstewart]

In that case I think we should add a ChangeDetectionTask and unstack CD datasets.

[robmarkcole]

MVP will focus on binary change detection, and report metrics per class change/no-change

[isaaccorley]

You could make a wrapper for TinyCD which would unstack the images in the forward pass.

[isaaccorley]

The downside of moving to keys of image1 and image2 instead of the original stacked channels format is that we can no longer use the datasets with SemanticSegmentationTask. So either way a preprocessing transform would need to be defined to restack the images together vs using image1/image2 keys

[robmarkcole]

Explored a solution to this in #1711

Additionally we could provide a custom transform to perform the stacking