SSMB: Self-Supervised Keypoint Detection Directly on Motion-Blurred Images

zhenjun_zhao · x · 2026-08-28

New paper SSMB introduces a self-supervised keypoint detector that works directly on motion-blurred images—no deblurring preprocessing, no handcrafted detector, no external pseudo-labels.

Key component: a Local Discriminability Enhancement (LDE) module restores fine local structure after global feature mixing. Training runs in two self-supervised stages: geometric pretraining on synthetic shapes, then blur-aware training on real sharp/blurred pairs. Code, models and datasets will be released upon acceptance.

Original post →

More from Research

Research channel →