GSPan: A Continuous Gaussian Primitive Representation for Arbitrary-Scale Pansharpening

2026-06-16 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

GSPan is a novel pansharpening framework that addresses the scale adaptation limitations of existing deep learning methods by introducing 2D Gaussian Splatting (GS). Instead of predicting pixels on a fixed grid, GSPan represents band-wise residual details as continuous, learnable 2D Gaussian primitives. Its Dual-Stream Hierarchical Interaction (DSHI) architecture, featuring a Spatial-Spectral Interactive Attention (SSIA) module, estimates these primitives from panchromatic (PAN) and multispectral (MS) observations. The predicted primitives are then rendered as a residual detail field and integrated into the upsampled MS image. This continuous representation enables GSPan to render fused images on arbitrary target sampling grids without requiring scale-specific retraining. Furthermore, it facilitates a Scale-Decoupled Asymmetric Inference (SDAI) strategy, which estimates primitives at a reduced resolution for efficient large-scene pansharpening. Experiments on QuickBird, GaoFen-2, WorldView-3, and WorldView-3-4K datasets demonstrate GSPan's leading fusion performance, with SDAI significantly accelerating inference while maintaining quality.

Key takeaway

For Computer Vision Engineers developing pansharpening solutions, GSPan offers a significant architectural shift. If you are struggling with fixed-grid limitations or computational costs for large scenes, consider adopting continuous Gaussian primitive representations. This approach allows arbitrary-scale rendering without retraining and enables efficient Scale-Decoupled Asymmetric Inference. Evaluate GSPan's methodology to enhance both the flexibility and performance of your image fusion pipelines.

Key insights

GSPan uses continuous Gaussian primitives for scale-adaptive pansharpening, overcoming fixed-grid limitations.

Principles

• Continuous representations enable arbitrary-scale image rendering.
• Decoupling primitive estimation from rendering boosts efficiency.
• Gaussian Splatting effectively models image residual details.

Method

Estimate band-wise residual details as 2D Gaussian primitives via a DSHI architecture with SSIA. Render these primitives into a residual field, then inject into the upsampled MS image.

In practice

• Explore 2D Gaussian Splatting for continuous image representations.
• Implement Scale-Decoupled Asymmetric Inference for large-scene processing.
• Integrate DSHI and SSIA for robust multispectral fusion.

Topics

• Pansharpening
• Gaussian Splatting
• Continuous Representation
• Scale-Decoupled Inference
• Image Fusion
• Remote Sensing Imagery

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Local-GS: Accelerating 3D Gaussian Splatting via Tile-Local Warp Coherence — Local-GS enhances 3DGS rendering by aligning Gaussian processing with GPU warp execution for improved efficiency.
• PanoPlane: Plane-Aware Panoramic Completion for Sparse-View Indoor 3D Gaussian Splatting — Panoramic scene completion with layout-anchored attention steering significantly improves sparse-view 3D reconstruction.
• AIGS-Net: Compact Illumination Field Modeling via 2D Gaussian Splatting for Fast Low-Light Image Enhancement — AIGS-Net uses adaptive 2D Gaussian Splatting and zero-parameter encoding for ultra-lightweight, fast low-light image enhancement.
• Where, What, Why: Toward Explainable 3D-GS Watermarking — A representation-native 3D-GS watermarking framework decouples carrier selection and visual compensation for robust, imperceptible, and explainable embedding.
• GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens — GlobalSplat uses global latent scene tokens for efficient, compact, and consistent 3D Gaussian Splatting.
• Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting — LGTM enables 4K novel view synthesis by decoupling geometry from resolution using textured Gaussian primitives.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.