While global deblurring (correcting for camera shake) has seen significant progress, local motion deblurring—where moving objects are blurred against a perfectly sharp background—remains a major challenge. This paper introduces ReLoBlur, the first large-scale real-world dataset for local motion blur, and LBAG, a gated neural network designed to localize and restore blurred regions without introducing artifacts into the static background.
The Problem: Local vs. Global Blur
Most deblurring research focuses on "global" blur, where the entire image is smeared due to camera movement. However, in scenarios like static security cameras or tripod-mounted photography, the background is sharp, but moving objects (cars, people, pets) are blurred.
Existing methods struggle with this for three primary reasons:
- Lack of Real-World Data: Synthetic datasets often fail to capture the complex physics of real-world light accumulation.
- Data Imbalance: Blurred objects typically occupy only a small fraction of the image (averaging ~12% in the ReLoBlur dataset). Standard networks waste computational resources processing already-sharp pixels.
- Abrupt Boundaries: Local blur ends abruptly at the edge of a moving object, a transition that global deblurring models are not tuned to handle.
The ReLoBlur Dataset Acquisition
To capture authentic pairs of blurred and sharp images, the authors constructed a custom synchronized beam-splitting system. This hardware setup allows a single scene to be captured simultaneously by two different sensors.
As shown in the figure above, the system splits the incoming light path:
- Camera B (Blur): Uses a long exposure to capture natural motion blur.
- Camera S (Sharp): Uses a very short exposure to capture the ground truth.
To ensure the images are usable for training, the team developed a sophisticated post-processing pipeline to correct for hardware-induced discrepancies.
This pipeline addresses color casts, brightness concordance, and sub-pixel misalignments, resulting in a final dataset of 2,405 high-quality real-world image pairs.
LBAG: Local Blur-Aware Gated Network
To solve the spatial imbalance of local blur, the authors introduced the Local Blur-Aware Gated (LBAG) network. Building on the MIMO-Unet architecture, LBAG utilizes specialized Gate Blocks that perform two tasks simultaneously:
- Localization: The network predicts a pixel-level "blur mask" to identify exactly where the motion blur is located.
- Gated Restoration: Using this mask, the network focuses its computational effort only on the blurred pixels, effectively "ignoring" the sharp background to prevent the introduction of unwanted artifacts.
Additionally, the authors implemented Blur-Aware Patch Cropping (BAPC). During training, this strategy ensures that at least 50% of the training patches contain blurred regions, forcing the model to learn from the most difficult parts of the image rather than the empty background.
Key Results and Insights
The proposed LBAG+ (pre-trained on the GoPro dataset) outperformed state-of-the-art global deblurring models across all metrics on the ReLoBlur dataset:
- PSNR: Achieved 34.85 dB, outperforming the nearest competitor (MIMO-Unet) by 0.33 dB.
- Weighted PSNR ($PSNR_w$): This metric focuses specifically on the blurred regions. LBAG+ scored 28.32, significantly higher than standard models, proving its superior ability to restore moving objects.
- Qualitative Performance: Visual results demonstrate that LBAG preserves fine textures (such as fabric or skin) that global models typically smear or distort.
Conclusion
By providing both a high-fidelity real-world dataset and a specialized gated architecture, this work addresses the "fidelity gap" in local motion deblurring. The gated mechanism ensures that restoration is applied only where needed, preserving the integrity of sharp backgrounds while successfully recovering details in moving subjects.
Summary updated