InternImageModelSize

Enums · AiDotNet.Enums

Defines the model size variants for InternImage (DCNv3-based CNN backbone).

For Beginners

InternImage comes in five sizes (Tiny through Huge). It proves that CNNs can match or exceed Vision Transformers when using modern deformable convolutions. Tiny is great for quick experiments, while XL and Huge target maximum accuracy.

How It Works

Technical Details: Each size uses a different DCNv3 backbone configuration with varying channel widths, depths, and group sizes. All variants use the UPerNet decoder.

Reference: Wang et al., “InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions”, CVPR 2023.

Fields

Field Summary
InternImageModelSize Base InternImage-B: Base variant (97M params).
InternImageModelSize Huge InternImage-H: Huge variant (1.08B params).
InternImageModelSize Small InternImage-S: Small variant (50M params).
InternImageModelSize Tiny InternImage-T: Tiny variant (30M params).
InternImageModelSize XL InternImage-XL: Extra-large variant (335M params).
int value__