Benchmarking 15 ways to reconstruct hyperspectral snapshots

What it does MST is a benchmarking toolbox and baseline collection for spectral compressive imaging reconstruction. It bundles over 15 algorithms—from recent learning-based transformers to classical model-based methods like TwIST and DeSCI—into a single comparison framework with pre-trained models, parameter counts, and FLOPS.

The interesting bit The authors treat spectral reconstruction as a competitive sport: their own MST++ won the NTIRE 2022 challenge, and the repo continuously absorbs new top-tier conference methods to keep the leaderboard honest. It is essentially high-quality glue code that saves you from rebuilding a dozen disparate academic repositories.

Key highlights

• 12 learning-based methods (CVPR/ECCV/NeurIPS) plus 3 classical model-based baselines
• MST++ won the NTIRE 2022 Spectral Reconstruction Challenge
• Pre-trained model zoo with quantitative comparisons (PSNR, SSIM, params, FLOPS)
• Supports both simulation datasets and real-scene experiments
• Includes binarized variants (BiSRNet) for resource-constrained reconstruction

Verdict Worth bookmarking if you work on snapshot compressive imaging or need to benchmark against recent top-tier methods. Skip it if you want a minimal single-algorithm pipeline—this is a broad survey collected under one roof.