MIORe: Motion Restoration Benchmark

• MIORe is a high-frame-rate dataset using adaptive frame averaging to synthesize realistic moderate motion blur based on optical flow statistics.
• It provides a multi-task benchmark with paired sharp and blurry frames, supporting deblurring, video frame interpolation, and optical flow estimation.
• The dataset challenges standard benchmarks by combining real high-FPS footage with synthetic blur generation to expose nuanced trade-offs in restoration tasks.

MIORe is a newly introduced dataset for motion restoration that was placed at the center of the AIM 2025 High FPS Non-Uniform Motion Deblurring Challenge and was subsequently presented, together with VAR-MIORe, as a multi-task benchmark for deblurring, video frame interpolation, and optical flow estimation. It is designed around high-frame-rate acquisition at 1000 FPS, full-HD imagery, and adaptive frame averaging driven by optical flow statistics in order to produce challenging yet regularized motion blur from real high-frame-rate footage. Across the two papers, MIORe is characterized as a benchmark that targets moderate real-world motion patterns, including non-uniform motion and occasional defocus blur, while VAR-MIORe extends the same acquisition pipeline to a broader motion-amplitude regime from minimal to extreme blur (Ciubotariu et al., 8 Sep 2025, Ciubotariu et al., 8 Sep 2025).

1. Definition and benchmark scope

MIORe was introduced to address what the dataset paper identifies as critical limitations in existing motion restoration benchmarks, including restricted motion diversity, simplistic or synthetic motion blur, limited environmental variety, and the lack of explicit control over motion amplitude (Ciubotariu et al., 8 Sep 2025). In the challenge paper, it is described as a newly introduced dataset created specifically to push the boundaries of single-image motion deblurring research by providing challenging, realistic benchmarks built from representative visual cues for complex aggregations of motion types (Ciubotariu et al., 8 Sep 2025).

The dataset is explicitly positioned as a benchmark for moderate blur. Its complementary dataset, VAR-MIORe, is described in two consistent but differently framed ways across the sources: in the challenge paper, it is said to push for extreme motion blur, up to four times the blur strength of MIORe; in the dataset paper, it is defined more broadly as spanning a variable range of motion magnitudes from minimal to extreme, with explicit control over motion amplitude (Ciubotariu et al., 8 Sep 2025, Ciubotariu et al., 8 Sep 2025). This pairing makes MIORe the regularized benchmark and VAR-MIORe the variable-amplitude extension.

A central design feature is that blurry images are synthetically generated from sharp high-frame-rate footage captured using a CHRONOS 2.1-HD camera, enabling precise ground-truth high-quality–input low-quality pairings across diverse real-life scenes (Ciubotariu et al., 8 Sep 2025). A common misconception is that “synthetic blur” necessarily denotes unrealistic image formation. In this case, the blur is synthetic in the sense of being generated by temporal averaging, but it is rooted in real 1000-FPS image sequences acquired under real scene dynamics and professional optics (Ciubotariu et al., 8 Sep 2025).

2. Acquisition pipeline and data construction

The acquisition protocol uses an industrial-grade CHRONOS 2.1-HD high-speed camera at full HD resolution, $1920 \times 1080$, and 1000 FPS, capturing over 1.1 million frames across 333 diverse sequences (Ciubotariu et al., 8 Sep 2025). The optical setup employs four professional lenses—Tamron 15–30mm, Canon 24mm, Sigma 85mm, and Laowa 100mm—providing wide-angle, standard, telephoto, and macro perspectives, with apertures adjusted for optimal light intake and varied for exposure matching (Ciubotariu et al., 8 Sep 2025).

The recorded sequences span both controlled and in-the-wild scenarios across seasons, weather, and lighting, including rain, fog, snow, and day-night cycles (Ciubotariu et al., 8 Sep 2025). The captured motion types include ego-camera translations, smooth or chaotic yaw, pitch, and roll rotations, multi-subject scene motion, dynamic backgrounds, fluid flows, parallax effects, and depth-dependent blur and defocus (Ciubotariu et al., 8 Sep 2025). The challenge paper independently emphasizes a wide range of moderate real-world motion patterns, not just translational motion but also complex and non-uniform motions, with occasional defocus blur to simulate real photography conditions more closely (Ciubotariu et al., 8 Sep 2025).

Blur synthesis is based on adaptive frame averaging driven by optical flow statistics rather than by averaging a fixed number of frames. The dataset paper gives two explicit examples: for scenes with higher displacement per frame, such as 10 px, only 3 frames are averaged; for low-motion scenes with subpixel displacement per frame, up to 30 frames are averaged (Ciubotariu et al., 8 Sep 2025). The stated objective is to ensure a normalized, consistent blur magnitude across the dataset for fair benchmarking.

The same paper reports the post-curation statistics of the two datasets: approximately 52,000 samples for MIORe and approximately 83,000 for VAR-MIORe, derived from the 333 source sequences (Ciubotariu et al., 8 Sep 2025). It also specifies the integration-window ranges used for blur generation: 3–35 frames averaged per blurry image for MIORe and 1–249 frames per blurry image for VAR-MIORe (Ciubotariu et al., 8 Sep 2025).

3. Tasks, annotations, and data organization

MIORe and VAR-MIORe are presented as multi-task datasets rather than as deblurring-only resources (Ciubotariu et al., 8 Sep 2025). For each blur instance, the middle frame is preserved sharp and serves as the deblurring and interpolation ground truth, while the sequence endpoints are also preserved sharp for video frame interpolation and optical flow estimation (Ciubotariu et al., 8 Sep 2025). Optical flow ground truth is provided as pseudo-labels, computed or enhanced using modern optical flow estimators such as DEQ-Flow (Ciubotariu et al., 8 Sep 2025).

The challenge paper defines a particular benchmark configuration for single-image motion deblurring. In Track 1, which uses MIORe, the training split contains 7,860 paired full-HD samples, the validation split contains 80 images with no ground truth shared, and the blind test split contains 75 images (Ciubotariu et al., 8 Sep 2025). In Track 2, which uses VAR-MIORe, the training split contains 1,299 paired samples, the validation split 67 images, and the blind test split 63 images (Ciubotariu et al., 8 Sep 2025). Ground truth is provided only for the training splits; validation and testing are blind, and outputs are evaluated on a central server (Ciubotariu et al., 8 Sep 2025).

The dataset paper reports broader corpus-level properties that are distinct from the challenge-track splits. These include 333 sequences, full-HD resolution, four lenses, and motion coverage summarized as “All” for MIORe in ego-camera translation and rotation, tracking, flat/parallax, and deformable content (Ciubotariu et al., 8 Sep 2025). In the comparison table reproduced there, MIORe is listed with maximum optical flow of 95, while VAR-MIORe is listed with maximum optical flow of 1932 px (Ciubotariu et al., 8 Sep 2025).

Attribute	MIORe	VAR-MIORe
Resolution	$1920 \times 1080$	$1920 \times 1080$
Acquisition FPS	1000	1000
Sequences	333	333
Blur window	3–35 frames	1–249 frames
Post-curation size	$\sim 52{,}000$	$\sim 83{,}000$
Max OF	95	1932

This organization supports two distinct uses. First, it provides a fixed experimental protocol for challenge-style benchmarking on moderate and extreme blur. Second, it provides a larger, multi-task corpus in which motion amplitude can be regularized or explicitly swept.

4. Evaluation protocol and metrics

Within the AIM 2025 challenge, MIORe functions as both the training source and the testbed for real-world effectiveness in Track 1 (Ciubotariu et al., 8 Sep 2025). The evaluation uses three metrics: PSNR, SSIM, and LPIPS, with the final rank defined as the average of metric-specific rankings rather than as the rank under a single scalar score (Ciubotariu et al., 8 Sep 2025). That ranking rule is methodologically important because it balances fidelity-oriented and perceptual criteria.

The challenge paper gives the PSNR definition as

$$\text{PSNR} = 10 \cdot \log_{10} \left( \frac{MAX^2}{\text{MSE}} \right),$$

where $\text{MSE}$ is the mean squared error between the restored and ground-truth images and $MAX$ is the maximum pixel value, 255 for 8-bit images (Ciubotariu et al., 8 Sep 2025).

It gives the SSIM definition as

$$\mathrm{SSIM}(x, y) = \frac{(2\mu_x\mu_y + C_1)(2\sigma_{xy} + C_2)}{(\mu_x^2+\mu_y^2+C_1)(\sigma_x^2+\sigma_y^2+C_2)},$$

and describes LPIPS as a deep-learning-based perceptual similarity metric, lower is better, using features from ImageNet-pretrained AlexNet (Ciubotariu et al., 8 Sep 2025).

The dataset paper additionally states that deblurring evaluation uses PSNR and SSIM for state-of-the-art methods such as AdaRevD, LoFormer, FFTFormer, UFPNet, and NAFNet; video frame interpolation evaluation reports PSNR and SSIM across motion groups and frame offsets; and optical-flow evaluation uses EPE and F1-score (Ciubotariu et al., 8 Sep 2025). It also notes that scenes can be bucketed by mean optical flow into motion-intensity groups such as Extremely Fast, Fast, and Medium-Fast, enabling per-motion-range performance analysis (Ciubotariu et al., 8 Sep 2025).

A common misunderstanding is to treat the challenge benchmark and the full dataset as identical entities. The papers separate them clearly: the challenge uses specific blind splits for deblurring, while the dataset itself is multi-task and supports controlled analyses over motion range and task type.

5. Challenge methods and empirical outcomes on MIORe

The AIM 2025 challenge registered 68 participants, with 9 teams ultimately submitting valid entries (Ciubotariu et al., 8 Sep 2025). On Track 1, which uses MIORe, Team VPEG obtained the top PSNR and SSIM, with 34.484 dB and 0.9026 respectively, and achieved final rank 1 (Ciubotariu et al., 8 Sep 2025). Team VPEG_2 followed with 34.155 PSNR, 0.8990 SSIM, and final rank 2, while BlurKing placed third with 33.337 PSNR and 0.8870 SSIM (Ciubotariu et al., 8 Sep 2025). Mier obtained the best LPIPS score, 0.1295, but ranked sixth overall because its PSNR and SSIM were lower, at 28.807 and 0.7800 (Ciubotariu et al., 8 Sep 2025).

These results are explicitly linked to different architectural choices. VPEG used Efficient Visual State Space Models and a multi-stage training protocol that started with GoPro and then focused specifically on MIORe, with progressive patch-size increase (Ciubotariu et al., 8 Sep 2025). VPEG_2 and VPEG_3 explored frequency-domain transformers, FFTformer+, benefiting from frequency attention and efficient aggregation tailored to MIORe’s moderately complex motion patterns (Ciubotariu et al., 8 Sep 2025). SRC-B employed a two-stage visual prompt-based system, directly trained and fine-tuned on MIORe; X-L combined EVSSM with multi-scale fusion; and Mier used a massive generative model trained on MIORe, emphasizing perceptual similarity in visually complex cases (Ciubotariu et al., 8 Sep 2025).

The challenge paper emphasizes several dataset-specific findings. It states that MIORe enabled a more nuanced and robust evaluation than legacy datasets such as GoPro, surfacing subtle distinctions in network capacity (Ciubotariu et al., 8 Sep 2025). It also reports that MIORe’s non-uniform, mild-to-moderate blur and hybrid motion-plus-defocus cases motivated architectural innovations including frequency-domain filtering, visual prompts, and multiscale interactions (Ciubotariu et al., 8 Sep 2025). The contrast between VPEG and Mier further illustrates the tension between perceptual and pixel-wise optimization: a model can excel on LPIPS while lagging on PSNR and SSIM (Ciubotariu et al., 8 Sep 2025). This suggests that MIORe is effective at exposing objective-function trade-offs that are less visible on simpler benchmarks.

6. Relation to prior benchmarks and research significance

The dataset paper contrasts MIORe and VAR-MIORe with GoPro, RealBlur, SINTEL, KITTI, and Vimeo90K, arguing that those benchmarks often use fixed frame rates, limited scene or equipment variety, narrow motion ranges, and blur synthesis that neglects the interplay of motion blur, defocus, and depth-variant effects (Ciubotariu et al., 8 Sep 2025). In the comparison table, MIORe is reported with higher spatial resolution than most prior benchmarks, four lenses rather than one, and broader motion-scenario coverage (Ciubotariu et al., 8 Sep 2025).

Several implications follow directly from the papers. First, MIORe provides regularized, motion-normalized blur through optical-flow-guided averaging, which the dataset paper frames as supporting fair benchmarking (Ciubotariu et al., 8 Sep 2025). Second, the challenge results indicate that using MIORe as a later stage in a curriculum after pretraining on GoPro can improve specialization to more realistic motion patterns (Ciubotariu et al., 8 Sep 2025). Third, because sharp left, middle, and right frames are preserved and optical-flow pseudo-labels are provided, the benchmark supports joint study of restoration tasks that are often evaluated in isolation (Ciubotariu et al., 8 Sep 2025).

The papers also identify realism-enhancing factors beyond motion alone: depth-dependent blur, camera focus effects, chromatic aberration, vignetting, adverse weather, seasonal transitions, and illumination variation (Ciubotariu et al., 8 Sep 2025). This suggests that the benchmark is designed not merely to increase difficulty, but to alter the structure of the evaluation problem toward camera-and-scene realism.

The project resources are listed at the MIORe GitHub page: https://github.com/george200150/MIORe (Ciubotariu et al., 8 Sep 2025, Ciubotariu et al., 8 Sep 2025).

7. Interpretive issues and prospective role

MIORe is sometimes read narrowly as a deblurring dataset because of its role in the AIM 2025 challenge. The dataset paper makes clear, however, that it is a multi-task benchmark with deblurring, video frame interpolation, and optical flow estimation all built into the data design (Ciubotariu et al., 8 Sep 2025). Conversely, it would also be incomplete to treat MIORe and VAR-MIORe as interchangeable. MIORe is the regularized moderate-blur benchmark; VAR-MIORe is the variable-range extension, ranging from 1 frame, effectively no blur, to 249 or more frames for extreme motion blur (Ciubotariu et al., 8 Sep 2025).

Another interpretive issue concerns “realism.” The dataset does not provide directly captured long-exposure blur from conventional cameras; instead, it synthesizes blur by averaging high-frame-rate real footage. The papers present this as a deliberate compromise: synthetic generation preserves exact sharp ground truth while the underlying scene dynamics, optics, and motion patterns remain real (Ciubotariu et al., 8 Sep 2025, Ciubotariu et al., 8 Sep 2025).

The challenge paper concludes that MIORe raised the difficulty bar, enabled fair and comprehensive evaluation through strong metrics and blind splits, and guided innovation in architecture, data handling, and evaluation (Ciubotariu et al., 8 Sep 2025). The dataset paper further characterizes MIORe and VAR-MIORe as the first datasets to capture the full real-world spectrum of motion blur under professional imaging conditions, provide scalable ground truths for multiple restoration tasks, and expose motion magnitude as an explicit evaluation parameter (Ciubotariu et al., 8 Sep 2025). A plausible implication is that MIORe’s enduring significance will depend less on its use as a single leaderboard benchmark than on its utility as an experimental substrate for motion-aware, multi-task restoration research.