SAMHQModelSize
Enums · AiDotNet.Enums
Defines the backbone size variants for SAM-HQ (High-Quality Segment Anything Model).
For Beginners
SAM-HQ extends the Segment Anything Model (SAM) with a High-Quality output token that produces significantly sharper and more accurate mask boundaries. It uses the same ViT backbone sizes as the original SAM.
How It Works
Technical Details: SAM-HQ adds an HQ output token and learnable global-local feature fusion to the original SAM architecture. The backbone is a Vision Transformer (ViT) with varying sizes. Training uses only 44K fine-grained masks.
Reference: Ke et al., “Segment Anything in High Quality”, NeurIPS 2023.
Fields
| Field | Summary |
|---|---|
SAMHQModelSize ViTBase |
ViT-B backbone (91M params). |
SAMHQModelSize ViTHuge |
ViT-H backbone (636M params). |
SAMHQModelSize ViTLarge |
ViT-L backbone (308M params). |
int value__ |