ObjectDetectionTask support for MSI

[adamjstewart]

Summary

The ObjectDetectionTask has an in_channels parameter but it isn't actually used for anything. At the moment, it seems that the trainer only supports RGB imagery. We should fix this.

Rationale

We specialize in MSI, how can we not support MSI.

Implementation

We're currently using torchvision backbones, which makes things more challenging. How hard would it be to switch to timm backbones?

Alternatives

Alternatively, we'll have to override the first channel of the torchvision backbone.

Additional information

No response

[robmarkcole]

The adaption for multi-channel looks straightforward - not sure about handling the pretrained weights

https://github.com/allenai/vessel-detection-sentinels/blob/main/src/models/frcnn.py