RSBlur: Real-World Deblurring Dataset

• RSBlur is a real-world image dataset pairing blurred images with sharp frame sequences to capture authentic sensor noise, motion blur, and ISP effects.
• The dataset uses a dual-camera system and a detailed synthesis pipeline to simulate realistic noise, saturation, and blur formation processes.
• RSBlur serves as a benchmarking tool for deblurring models, enhancing generalization and guiding efficient restoration methods for real-world applications.

The RSBlur dataset is a real-world image dataset designed to address shortcomings in synthetic blur formation for learning-based image deblurring. It consists of paired real blurred images and corresponding sharp image sequences, enabling a systematic analysis of physical blur characteristics and the discrepancies between real and synthetic blurred imagery. Its construction and associated analysis pipeline have established RSBlur as a benchmark for both dataset realism and the design of synthesis methods that better reflect sensor, motion, and image signal processing (ISP) effects observed in practice (Rim et al., 2022). RSBlur has since become central to image restoration benchmarking, model evaluation, and challenge competitions emphasizing efficient, real-world deblurring (Feijoo et al., 14 Oct 2025).

1. Dataset Construction and Composition

RSBlur comprises 13,358 real blurred images from 697 outdoor scenes, with each blurred image paired with a sequence of nine sharp images recorded during the corresponding exposure. Acquisition employs a dual-camera system: one module with a 5% neutral density filter captures a long-exposure blurred image ($0.1\,\mathrm{s}$), while the other records nine consecutive sharp images at $0.005\,\mathrm{s}$ per frame. Images are initially recorded in RAW format before conversion to sRGB space via a pipeline that includes white balancing, demosaicing, color correction, and gamma correction performed using a camera ISP.

All scenes in RSBlur reflect natural motion blurs from handheld operation, capturing both camera shake and object motion. The real blurred images retain characteristics absent from most synthetic datasets—discontinuity, sensor noise, and saturation—making the dataset representative of authentic photographic conditions.

The data split for machine learning competitions utilizing RSBlur is as follows (Feijoo et al., 14 Oct 2025):

Set	Paired Images	Scenes
Train	8,887
Validation	1,120
Test	3,360
Final Test	420	84 (disjoint)

2. Physical and Algorithmic Analysis of Blur

Unlike conventional synthetic datasets that generate blur by frame averaging and interpolation (e.g., GoPro, REDS), RSBlur highlights several critical factors differentiating real and synthetic blur:

• Discontinuities in Trajectories: Synthetic methods, even with frame interpolation, cannot perfectly mimic the continuous, accumulative light integration process of camera sensors.
• Saturated Pixels: Sensor saturation from limited dynamic range creates distinct blur artifacts absent in averaged synthetic data.
• Noise Properties: Real blurred images from long exposures include photon shot noise, read noise, and quantization noise, which synthetic averaging reduces or omits.
• ISP Effects: Real blur undergoes nonlinear and sensor-dependent transformations not captured in naive synthetic pipelines.

Direct comparison of the paired real blurred image and its sharp sequence in RSBlur enables granular examination and synthesis of these physical phenomena, providing a reference for realistic data generation and restoration benchmarking.

3. Blur Synthesis Pipeline

Building on analysis enabled by RSBlur, a comprehensive blur synthesis pipeline was proposed to simulate real blur with high fidelity (Rim et al., 2022):

• Frame Interpolation: Sharp sequence ($N=9$) interpolated to $N=65$ frames (ABME method) to fill temporal gaps.
• Linearization and Averaging: Frames converted from sRGB to linear space via gamma decoding; linear images are averaged to simulate sensor integration over time.
• Saturation Synthesis:

For sharp image $S_i$, mask $M_i(x,y,c) = 1$ if $S_i(x,y,c)=1$, else $0$. Aggregate: $M_\mathrm{sat} = \frac{1}{N} \sum_i M_i$. Final synthesis: $$B_\mathrm{sat} = \mathrm{clip}(B_\mathrm{syn} + \alpha M_\mathrm{sat})$$, where $\alpha \sim \mathcal{U}(0.25, 1.75)$, $\mathrm{clip}(\cdot)$ limits values to $[0,1]$.

• RAW Conversion: Reverse ISP to the camera RAW space using inverse color correction, mosaicing, and inverse white balance.
• Noise Modeling: Add shot and read noise in RAW:

$$B_\mathrm{noisy} = \beta_1 (I + N_\mathrm{shot}) + N_\mathrm{read}$$

with $$N_\mathrm{shot} \sim \mathrm{Poisson}(\beta_1 B_\mathrm{raw})$$, $N_\mathrm{read} \sim \mathcal{N}(0, \beta_2)$.

• ISP Application: Reapply ISP to obtain final sRGB output.

This pipeline explicitly models saturation, realistic noise, and sensor-specific ISP transformations, generating synthetic blur nearly indistinguishable—statistically and visually—from real blur in RSBlur.

4. Benchmarking and Model Evaluation

Extensive experiments have demonstrated that deblurring models trained with images synthesized via the RSBlur pipeline outperform those trained on conventional synthetic data. For instance, SRN-DeblurNet trained with this pipeline achieved PSNR scores (e.g., $32.06\,\mathrm{dB}$) close to the best results from real camera-captured data. Inclusion of saturation masks and physical noise models proved essential for closing the “domain gap.”

Recent competitions, such as the AIM 2025 Efficient Real-World Deblurring Challenge (Feijoo et al., 14 Oct 2025), leverage RSBlur as the core testbed. Models are evaluated using a composite metric:

$$\mathrm{Score} = \lambda_1 \cdot \mathrm{PSNR} + \lambda_2 \cdot \mathrm{SSIM} + \lambda_3 \cdot \mathrm{LPIPS},$$

where $\lambda_1, \lambda_2, \lambda_3$ are weighting factors for fidelity and perceptual quality. Strict computational constraints are imposed ($<5$M parameters, $<200$ GMACs). State-of-the-art solutions based on transformer architectures and efficient attention mechanisms (e.g., NAFRepLocal, RestormerL) have demonstrated performance at $31.13\,\mathrm{dB}$ PSNR.

5. Impact on Benchmarking, Generalization, and Methodology

RSBlur and its synthesis protocol have led to key advancements:

• Dataset Realism: Inclusion of authentic sensor phenomena in training data leads to substantial improvements in model generalization and robustness to real-world blur.
• Comparative Studies: RSBlur allows side-by-side comparison against synthetic benchmarks, revealing substantial performance gaps for existing models and guiding future architecture design for deblurring.
• Cross-Dataset Transfer: Models trained with RSBlur-synthesized data exhibit superior generalization to other real blur datasets (e.g., RealBlur_J, BSD) compared to those trained purely on synthetic or lab-generated data.
• Challenge Design: The use of strict efficiency constraints driven by RSBlur procurement and benchmarking supports deployment-ready model development for mobile and edge devices.

6. Applications and Broader Significance

The RSBlur dataset is directly applicable to:

• Image Restoration and Deblurring: Training and evaluating single-image and video deblurring models with realistic noise, saturation, and blur artifacts.
• Super-Resolution and Quality Assessment: Serving as a data source for perceptual super-resolution (cf. ReBlurSR, PBaSR (Qin et al., 2024)) and blur assessment frameworks.
• Domain Adaptation and Transfer Learning: As a benchmark for transfer learning and domain adaptation studies targeting real-world blur restoration.
• Efficient Model Design: Providing the baseline for challenge-driven architecture optimization and efficiency research.

A plausible implication is that the RSBlur acquisition protocol and synthesis approach are being adopted as best practices for future real-world datasets, motivating a shift in benchmarking away from synthetic blur toward representative, physically accurate image data.

7. Limitations and Directions for Advancement

While RSBlur sets a high standard for data realism, limitations remain—in particular, broader scene and device diversity, low-light sensitivity, and spatially nonuniform blur dynamics typical of some photographic conditions are only partly addressed. Ongoing research seeks to augment the RSBlur protocol with smartphone-based slow-motion video datasets (Noki et al., 24 Jun 2025), richer noise characterizations, and metadata-driven domain adaptation.

RSBlur represents a foundational resource for the development, benchmarking, and practical deployment of robust image restoration technologies in realistic operational environments.