Mask2FormerModelSize

Enums · AiDotNet.Enums

Defines the backbone size variants for Mask2Former.

For Beginners

Mask2Former is a universal segmentation model that handles semantic, instance, and panoptic segmentation with a single architecture. The backbone size controls feature extraction capacity. Swin-T is efficient for prototyping, while Swin-L offers the highest accuracy.

How It Works

Technical Details: Mask2Former uses a Swin Transformer or ResNet backbone with a pixel decoder (Multi-Scale Deformable Attention) and a transformer decoder with masked cross-attention for query-based mask prediction.

Reference: Cheng et al., “Masked-attention Mask Transformer for Universal Image Segmentation”, CVPR 2022.

Fields

Field Summary
Mask2FormerModelSize R101 ResNet-101 backbone (63M params).
Mask2FormerModelSize R50 ResNet-50 backbone (44M params).
Mask2FormerModelSize SwinBase Swin-B backbone (107M params).
Mask2FormerModelSize SwinLarge Swin-L backbone (216M params).
Mask2FormerModelSize SwinSmall Swin-S backbone (69M params).
Mask2FormerModelSize SwinTiny Swin-T backbone (47M params).
int value__