SegNeXtModelSize

Enums · AiDotNet.Enums

Defines the model size variants for SegNeXt (Multi-Scale Convolutional Attention backbone).

For Beginners

SegNeXt comes in four sizes (Tiny through Large). Smaller sizes (Tiny) are faster and use less memory, while larger sizes (Large) are more accurate but require more compute. Tiny is great for real-time applications, while Base and Large offer excellent accuracy for production deployments.

How It Works

Technical Details: Each size uses a different MSCAN (Multi-Scale Convolutional Attention Network) backbone with varying channel widths and encoder depths. All variants use the Hamburger decoder for semantic segmentation.

Reference: Guo et al., “SegNeXt: Rethinking Convolutional Attention Design for Semantic Segmentation”, NeurIPS 2022.

Fields

Field Summary
SegNeXtModelSize Base MSCAN-B: Base variant (27.6M params).
SegNeXtModelSize Large MSCAN-L: Large variant (48.9M params).
SegNeXtModelSize Small MSCAN-S: Small variant (13.9M params).
SegNeXtModelSize Tiny MSCAN-T: Tiny variant (4.3M params).
int value__