Mamba state-space models tackle event-guided video deblurring at ICCV 2025

EVDM: Event-based Real-world Video Deblurring with Mamba (ICCV 2025) applies the recently-prominent Mamba state-space model architecture to event-guided video deblurring, a task that combines two challenging sub-problems: leveraging asynchronous event data to compensate for motion blur in a synchronized frame video stream.

The deblurring use case

A frame camera operating next to an event camera captures visual detail that the event camera cannot: color, absolute intensity, textures. But the frame camera blurs fast-moving objects across its integration window while the event camera captures the same motion with microsecond precision.

Event-guided deblurring uses the event camera’s motion measurement to deconvolve the blur from the frame image. Given a blurred frame and a record of which pixels changed brightness and when during the exposure window, the exposure trajectory of the camera can be reconstructed and inverted, producing a sharp image.

This is valuable in a dual-modal sensing configuration: keep the frame camera for scene reconstruction and classification tasks that benefit from dense color imagery, use the event camera to sharpen the output where motion blur degrades it.

Why Mamba over transformers

Previous event-guided deblurring approaches used convolutional or transformer architectures to fuse the event and frame streams. The Mamba state-space model architecture is better suited to this task for a specific reason: events are a sequence of discrete, variable-rate time-stamped measurements, and Mamba was designed for efficient long-sequence processing.

Transformers process all elements of a sequence in parallel, which requires storing and computing over the entire sequence. For event streams where the sequence length can reach millions of events per second, this is prohibitively expensive. Mamba processes sequences incrementally with fixed memory, handling variable-length event streams without memory scaling issues.

The paper demonstrates that EVDM achieves reconstruction quality comparable to transformer baselines at roughly 60% of the parameter count and significantly lower inference time.

Real-world performance

Most prior deblurring benchmarks use synthetic blur — frames with artificially lengthened virtual exposures. The paper evaluates on real-world blurred video from a handheld camera, where the blur is less regular and harder to deconvolve. Performance in this regime is generally lower than on synthetic benchmarks, and EVDM is no exception — but it maintains a meaningful advantage over frame-only deblurring approaches.

Implications for dual-modal sensing

The case for a dual-modal event-plus-frame sensor in a defense application comes down to whether the two sensors are better together than either alone. The event camera alone cannot capture the scene texture needed for target classification. The frame camera alone blurs at the speeds that matter for terminal guidance.

Event-guided deblurring is the glue layer: it takes the texture from the frame channel and the motion measurement from the event channel and produces a clean, high-speed image sequence that neither sensor could produce individually.

EVDM demonstrates that this fusion can run efficiently enough for embedded deployment. Whether the resulting image quality improvement is worth the additional sensor mass and processing overhead is an application-specific tradeoff — but the technical feasibility is established.

---

EVDM: Event-based Real-world Video Deblurring with Mamba. ICCV 2025.