RobustSCI: Beyond reconstruction to restoration for snapshot compressive imaging under real-world degradations

Deep learning has greatly advanced video Snapshot Compressive Imaging (SCI), but most existing methods are developed for reconstructing videos from relatively clean or idealized compressive measurements. In practical capture, motion, low illumination, and sensor imperfections can degrade the scene signal before and during compressive encoding. In such cases, faithfully reconstructing the captured signal is not sufficient when the desired output is the underlying clean scene. This paper studies degraded video SCI restoration, which aims to recover clean video frames directly from degraded compressive measurements. We construct a benchmark based on DAVIS 2017 by introducing controlled motion-blur, low-light, and mixed degradations before SCI measurement generation, and we clarify the physical approximations and limitations of this simulation. We propose RobustSCI, an encoder–decoder network equipped with degradation-aware RobustCFormer blocks that combine multi-scale deblurring and frequency enhancement. We further introduce RobustSCI-C, an optional cascaded variant that applies a separately trained deblurring network for severe motion blur, providing additional restoration quality at the cost of higher model complexity. Experiments on simulated degraded benchmarks show favorable performance against representative SCI reconstruction models. Experiments on real CACTI measurements provide complementary qualitative evidence beyond simulation; because ground-truth high-speed frames are unavailable for these captures, we interpret the real-data results as qualitative evidence rather than quantitative verification of pixel-level fidelity.