Researchers release largest real-world image deblurring dataset from smartphone slow-motion video

Computer vision researchers have unveiled a large-scale image deblurring dataset constructed from smartphone high-speed video, aiming to advance the field of motion blur removal and reveal gaps in current approaches.

The Dataset#

The benchmark contains more than 42,000 high-resolution pairs of blurry and sharp images—roughly 10 times larger than previous widely adopted datasets. The research team created the dataset by capturing sequences of 240 frames per second using smartphone slow-motion recording, then synthetically generating blur by averaging consecutive frames while retaining the temporally centered frame as the ground-truth sharp reference. The collection spans diverse real-world scenarios, including both indoor and outdoor settings with varying degrees of object and camera movement, offering 8 times more scene diversity than existing benchmarks.

Performance Gap#

When the researchers evaluated state-of-the-art deblurring models on their new dataset, they observed significant performance drops compared to results on older benchmarks. This degradation underscores the complexity and variety present in the new data, suggesting that existing methods may rely on patterns specific to synthetic or limited datasets rather than robust, generalizable deblurring strategies.

Why It Matters#

The dataset serves as a challenging evaluation tool designed to push toward more practical and reliable deblurring systems. By capturing authentic blur patterns from consumer-grade smartphone cameras operating at high frame rates, the benchmark reflects real conditions that practitioners encounter—a more accurate test of whether models can actually remove blur from photographs and video in practical applications.