NTIRE 2025 deblurring challenge reveals the state of event-guided image restoration

NTIRE 2025 Challenge on Event-Based Image Deblurring: Methods and Results (arxiv 2504.12401) documents the outcomes of a competitive challenge in which research teams submitted their best event-guided deblurring systems to be evaluated on a common held-out test set. Challenge papers of this type are valuable because they provide an honest, reproducible snapshot of where a field stands at a given moment.

What deblurring challenges measure

The task: given a blurred frame and a synchronized event stream that recorded what happened during the frame’s exposure window, recover a sharp image. The event stream provides ground truth about which pixels moved and by how much during exposure. The challenge is to use that information effectively to sharpen the image without introducing artifacts, ghosting, or color shifts.

Participants submit methods ranging from classic optimization approaches (formulate the deblurring as a regularized inverse problem) to fully learned approaches (train an end-to-end network on paired blurred-sharp data) to hybrid approaches (use events to initialize an optimization that a network then refines).

The 2025 results

The winning submissions in the 2025 challenge achieved PSNR scores consistently above 35 dB on the challenge test set — significantly higher than the baseline of 28 dB achieved by frame-only deblurring methods, demonstrating that event camera data genuinely helps.

The performance spread between participants was smaller than in prior challenges, suggesting that the method landscape is converging. The top five submissions performed within about 0.8 dB of each other, indicating that incremental architectural improvements are unlikely to produce large gains without better data or a different approach to the problem.

Where the hard cases remain

The challenge paper analyzes the failure cases for all top submissions. The common thread is targets in saturated image regions: when a target is imaged against a very bright sky or near a light source that saturates the sensor, events from that region are unreliable (saturated pixels cannot generate change events) and the deblurring fails. This is precisely the scenario that occurs with an airborne UAS against a high-sun sky.

This limitation is worth noting: event cameras are often cited as handling high dynamic range better than frame cameras, and they do at the sensor level — each pixel adapts its local operating point. But the deblurring task uses the frame camera as the primary image source, and that source saturates. The event camera provides motion information about non-saturated regions but cannot help where the frame saturates.

Hybrid event-frame cameras

One notable direction in the challenge submissions is processing from hybrid cameras — single sensors that output both events and frames from the same pixel array. Sony and other vendors have produced hybrid sensors that allow per-pixel switching between event and frame modes. On a hybrid sensor, the frame and event streams are perfectly co-registered (no parallax, no calibration uncertainty) which simplifies the fusion considerably.

For seeker applications, a single hybrid sensor is more mass-and-volume-efficient than two separate sensors and eliminates the calibration burden. This direction is receiving increasing research attention and is worth tracking.

---

NTIRE 2025 Challenge on Event-Based Image Deblurring: Methods and Results. arxiv 2504.12401