Using Mask R-CNN for Object Detection

[FogDrip]

I noticed through several articles that TorchGeo uses a Fast R-CNN with their ObjectDetection task. Is it possible to use something like Mask R-CNN with this task instead? I would really like to try out Mask R-CNN for my habitat mapping effort. Additionally, I'm also trying to get a good scope of all of the types of tasks and models compatible with TorchGeo, as I don't see all of them listed directly in the models or task pages on the Read the Docs website.

Thanks!

[adamjstewart]

I noticed through several articles that TorchGeo uses a Fast R-CNN with their ObjectDetection task.

We currently support Faster R-CNN, FCOS, and RetinaNet

Is it possible to use something like Mask R-CNN with this task instead?

Definitely! Torchvision has Mask R-CNN, so we just need to add it as an option in detection.py. Want to open a PR?

I'm also trying to get a good scope of all of the types of tasks and models compatible with TorchGeo, as I don't see all of them listed directly in the models or task pages on the Read the Docs website.

Types of tasks can be found here. Types of models are everything in timm (700+ models), SMP, some models from torchvision, and any models listed here.

Of course, you don't have to use our PyTorch Lightning trainers, you could use our data loaders and combine them with any model ever written, even ones for TensorFlow.