Resolution-Adaptive Networks

Resolution-Adaptive Networks are neural networks designed to operate effectively across a wide range of input resolutions — a single model handles inputs from low to high resolution, adapting its processing to the available resolution without requiring separate models for each resolution.

Resolution Adaptation Methods

• Multi-Scale Training: Train on inputs at various resolutions — model learns to handle any resolution.
• Resolution-Dependent Channels: Allocate more channels at higher resolutions for proportional compute scaling.
• Feature Pyramid Networks (FPN): Multi-resolution feature extraction with top-down and lateral connections.
• Resolution Policy: Lightweight module decides the optimal resolution for each input.

Why It Matters

• Flexible Input: Real-world inputs come at varying resolutions — sensors, cameras, and equipment produce different resolutions.
• Efficiency: Low-resolution inference for simple cases saves 4-16× computation (quadratic scaling).
• Quality Scaling: When more compute is available, process at higher resolution for better accuracy.

Resolution-Adaptive Networks are scale-agnostic models — handling any input resolution within a single network for flexible, efficient inference.