EoMTModelSize

Enums · AiDotNet.Enums

Defines the backbone size variants for EoMT (Encoder-only Mask Transformer).

For Beginners

EoMT removes the complex pixel decoder and transformer decoder used by models like Mask2Former, instead placing mask queries directly inside a plain Vision Transformer (ViT/DINOv2). This makes it 4.4x faster than Mask2Former while maintaining competitive accuracy.

How It Works

Technical Details: Uses DINOv2 as the backbone. Queries are inserted at intermediate ViT layers and processed alongside image tokens. No separate decoder needed. Achieves strong results on COCO panoptic, ADE20K semantic, and Cityscapes instance.

Reference: Saporta et al., “Encoder-only Mask Transformer”, CVPR 2025 Highlight.

Fields

Field Summary
EoMTModelSize Base ViT-B/DINOv2 backbone (~86M params).
EoMTModelSize Large ViT-L/DINOv2 backbone (~307M params).
EoMTModelSize Small ViT-S/DINOv2 backbone (~22M params).
int value__