Input-Aware Sparse Computation for Scalable Medical and Remote Sensing Vision Models

Abstract

Scalable vision models in medical imaging and remote sensing require computational infrastructure that can process large multimodal datasets efficiently. This topic centers on input-aware sparse matrix-matrix multiplication, hyperspectral-LiDAR fusion, forensic vision-language compression, and instance-aware super-resolution. Sparse computation is especially relevant for graph-based biomedical reasoning, large image feature maps, multimodal token representations, and cross-domain retrieval systems. Hyperspectral-LiDAR fusion introduces high-dimensional spatial-spectral data, while forensic vision-language models and super-resolution systems increase the need for efficient token and feature processing. The literature structure highlights the computational layer underlying reliable medical and geospatial AI, where efficiency, scalability, and representation fidelity must be balanced.