Realblur

Realblur is a term used in photography and computational imaging to describe motion blur that occurs in real-world scenes during exposure, as opposed to blur produced synthetically by applying a blur kernel to a sharp image. In research, RealBlur also refers to datasets consisting of real blurred images paired with corresponding sharp references, created to provide realistic benchmarks for deblurring algorithms.

The RealBlur datasets aim to capture the complexities of blur found in everyday photography, including camera motion, object motion, varying lighting, and cluttered scenes. RealBlur images are paired with sharp references that approximate the scene without blur, enabling researchers to train and evaluate deblurring models on data that better reflects real-world conditions. The datasets are typically collected from multiple camera types, including consumer cameras and smartphones, to cover a range of imaging characteristics and noise patterns.

Two widely used subsets of the RealBlur collection are RealBlur-J and RealBlur-R, which originate from different acquisition setups and device categories. These subsets help researchers assess an algorithm’s robustness to device-specific blur and to generalize across imaging pipelines. RealBlur is often used for blind deconvolution and deblurring tasks, with evaluation focusing on both objective metrics such as PSNR and SSIM and perceptual quality.

Impact and usage: RealBlur has become a standard resource for advancing real-world deblurring, encouraging models trained on realistic blur to generalize better to photographs taken in everyday conditions. The datasets are publicly released for research purposes and are commonly cited in studies comparing deblurring approaches. Related topics include motion blur, image restoration, and image quality assessment.