OneFormerModelSize

Enums · AiDotNet.Enums

Defines the backbone size variants for OneFormer.

For Beginners

OneFormer is a universal segmentation model that handles all three segmentation tasks (semantic, instance, panoptic) with a single model trained only on panoptic data. It uses text conditioning to switch between tasks at inference time.

How It Works

Technical Details: Built on top of Mask2Former with a text encoder that conditions the segmentation on a task description. Uses Swin or DiNAT (Dilated Neighborhood Attention Transformer) backbones.

Reference: Jain et al., “OneFormer: One Transformer to Rule Universal Image Segmentation”, CVPR 2023.

Fields

Field Summary
OneFormerModelSize DiNATLarge DiNAT-L backbone (223M params).
OneFormerModelSize SwinLarge Swin-L backbone (219M params).
int value__