Go to ICLR 2026 Conference homepageRED: Robust Event-Guided Motion Deblurring with Modality-Specific Disentangled Representation
Keywords: event camera, motion deblur, image deblur
Abstract: Event cameras provide sparse yet temporally high-resolution motion information, demonstrating great potential for motion deblurring.
However, the delicate events are highly susceptible to noise. Although noise can be reduced by raising the threshold of Dynamic Vision Sensors (DVS), this inevitably causes under-reporting of events. Most existing event-guided deblurring methods overlook this practical trade-off. The modality-indiscriminate feature extraction and naive fusion treat images and events as statistically similar inputs, leading to unstable and mixed representations, especially when events are disrupted. To tackle these challenges, we propose a Robust Event-guided Deblurring (RED) network with modality-specific disentangled representation. First, we introduce a Robustness-Oriented Perturbation Strategy (RPS) that mimics various DVS thresholds, exposing RED to diverse under-reporting patterns and thereby fostering robustness under unknown conditions. To better exploit partially disrupted events, we design a Modality-specific Representation Mechanism (MRM) that disentangles the inputs into three complementary components: an image-semantic representation capturing structure and textures, an event-motion representation extracting fine-grained motion details, and a cross-modal representation modeling complementary interactions. Building on these reliable features, two interactive modules are presented to enhance motion-sensitive areas in blurry images and inject semantic context into under-reporting event representations.
Extensive experiments on synthetic and real-world datasets demonstrate RED consistently achieves state-of-the-art performance in terms of both accuracy and robustness.
Primary Area: applications to computer vision, audio, language, and other modalities
Submission Number: 6537
Loading