Add Segmentation Models.

[oke-aditya]

🚀 Feature

This should take some study and time.
Supporting Torchvision Segmentation models should be simple.
But we would need to support backbones, etc.

A nice place to know about these models is here

These models are good, but again we need some decomposition and refactoring which would work better with our API.

Segmentation has following tasks.

1. Instance Segmentation.
2. Semantic Segmentation.
3. Panopatic Segmentation.

But these tasks are kind of superset of each other, should be decompose models to each part or keep unified to segmentation.

How do we structure backbones for such APIs.

Detectron2 is very good in handling such cases. We need to provide segmentation backbones.

Let's try once with torch models so we get some API design for now.

We can then iterate over it

Torchvision models to Support

• FCN
• DeepLabv3
• MaskRCNN

All 3 will have almost similar recipe etc.

Simpler models which we can add

• LinkNet
• U-Net