VAR-MIORe: A Multi-Task Motion Benchmark

• VAR-MIORe is a multi-task benchmark that evaluates restoration and motion-analysis techniques across a controlled range of motion amplitudes.
• It employs 1000 FPS, 1920×1080 video captured with diverse professional lenses and an adaptive frame averaging protocol to synthesize blur.
• Empirical results highlight that as motion magnitude increases, state-of-the-art deblurring, frame interpolation, and optical flow methods suffer significant performance degradation.

VAR-MIORe is a multi-task, high-frame-rate real-world benchmark designed to evaluate restoration and motion-analysis methods under a continuously variable range of motion magnitudes, from minimal or sharp-like motion to extreme motion. Built from 1000 FPS, 1920×1080 video captured with an industrial camera and multiple professional lenses, it extends MIORe by making motion amplitude itself an explicit benchmark variable for single-image motion deblurring, video frame interpolation (VFI), and optical flow (OF) estimation. Its central contribution is not merely additional data volume, but a protocol in which the same acquisition pipeline is used to sweep motion magnitude across a controlled range, thereby exposing how performance degrades as displacement and blur severity increase (Ciubotariu et al., 8 Sep 2025).

1. Benchmark definition and relation to MIORe

VAR-MIORe is the extension of MIORe that spans a variable range of motion magnitudes, whereas MIORe is a regularized multi-task benchmark in which blur magnitude is made relatively comparable across scenes by adaptively averaging a variable number of high-speed frames based on motion estimates (Ciubotariu et al., 8 Sep 2025). MIORe targets a mid-level blur regime, approximately centered around 30 pixels of optical flow, so that scenes with different true speeds produce roughly comparable blur severity. VAR-MIORe, by contrast, is positioned as a stress-test benchmark that probes the breaking points of current methods as motion magnitude increases from almost no motion to very large displacement or severe blur.

The design is motivated by limitations in prior restoration datasets. The paper identifies limited motion variability, insufficient real-world coupling between blur and scene dynamics, fixed or weak control over blur generation, and the absence of explicit motion-amplitude benchmarking. In that framing, VAR-MIORe is intended to answer questions such as how performance degrades as displacement grows, which methods fail first, and whether methods trained on moderate-motion datasets remain robust under large motion. The paper characterizes it as the first benchmark with explicit control over motion amplitude (Ciubotariu et al., 8 Sep 2025).

2. Acquisition pipeline, scene coverage, and corpus

The acquisition setup uses a CHRONOS 2.1-HD high-speed camera at 1920 × 1080 and 1000 FPS, with a per-recording limit of up to 5516 frames due to internal RAM limits (Ciubotariu et al., 8 Sep 2025). Four professional lenses are used: Tamron 15–30mm zoom, Canon 24mm, Sigma 85mm, and Laowa 100mm. This lens diversity is intended to introduce realistic optical variation, including lens degradation models, vignetting, chromatic aberration, and defocus or bokeh behavior. Lenses were used mostly wide open to maximize light, with narrower apertures in bright scenes when needed for exposure control.

The capture is primarily in-the-wild, spanning natural and urban scenes, all four seasons, different times of day, and conditions including clouds, fog, rain, snow, as well as internal adverse conditions such as defocus, overexposure, and underexposure. The selected corpus contains 333 meticulously selected sequences based on 1,147,507 raw frames. Scene selection is described as maximizing diversity in motion granularity, scene type, seasons, adverse conditions, and optics.

Motion scenarios include both ego-camera motion and scene/object motion. Ego-motion categories include static capture, translations along $O_x, O_y, O_z$, radial translation or zooming, dolly zoom, and yaw, pitch, roll rotations, as well as cinematic motions such as panning, tilt shots, and barrel roll. Scene motion includes independently moving vehicles, humans, animals, and, conceptually, liquids, fire, and smoke, together with both flat background motion and parallax or depth-layered motion. The paper also emphasizes dynamic multi-subject interactions and depth-dependent blur, so that motion blur, defocus blur, parallax, and occlusion can coexist in the same real captures (Ciubotariu et al., 8 Sep 2025).

3. Motion-amplitude construction and blur generation

VAR-MIORe operationalizes motion amplitude through optical flow magnitude and through the temporal integration span used to synthesize blur (Ciubotariu et al., 8 Sep 2025). The paper explicitly says that both mean optical flow and maximum optical flow are used adaptively. A faithful reconstruction of the flow magnitude used in the benchmark is

$$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$

with image-level summaries

$$\mu_f = \frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\mathbf{f}(\mathbf{x})\|_2, \qquad f_{\max} = \max_{\mathbf{x}\in\Omega}\|\mathbf{f}(\mathbf{x})\|_2.$$

Blur synthesis is based on averaging high-speed sharp frames while preserving sharp supervision frames. In symmetric form, a faithful reconstruction is

$B_t = \frac{1}{2K+1}\sum_{i=-K}^{K} I_{t+i},$

with the center sharp frame $I_t$ retained as ground truth and the endpoints $I_{t-K}$ and $I_{t+K}$ preserved as sharp inputs for VFI and OF. The exact adaptive sampling rule is not published, but the paper states that MIORe targets a mid-level blur around 30 pixels of optical flow, averaging fewer frames when motion is large and more frames when motion is small. A faithful rule template is

$N = \phi(\mu_f, f_{\max}),$

where $N$ decreases as motion magnitude increases.

The distinction between MIORe and VAR-MIORe is quantitative as well as conceptual. For MIORe, blur is controlled to stay in a comparable regime by varying the number of averaged sharp frames roughly between 3 and 35 composed sharp frames per blurry image; the table reports Offset = [3, 35] ms and Final FPS = [28, 1000]. For VAR-MIORe, the motion range is much wider: Offset = [1, 249] ms, with benchmark buckets corresponding to averaging 1, 5, 13, 29, 61, 125, 249 frames, Final FPS = [4, 1000], and Max OF up to 1932, versus 95 for MIORe. The paper presents these buckets as the clearest protocol for controlled motion-magnitude variation (Ciubotariu et al., 8 Sep 2025).

4. Supported tasks, supervision, and annotations

For every generated sample, the benchmark stores a left sharp frame, middle sharp frame, right sharp frame, and a blurred frame produced by averaging (Ciubotariu et al., 8 Sep 2025). This supports three supervised tasks with a single acquisition pipeline. In motion deblurring, the input is the blurred image and the supervision target is the sharp center frame. In video frame interpolation, the inputs are the left and right sharp frames and the target is again the center sharp frame. In optical flow estimation, the inputs are the left and right sharp frames, while the ground truth is a carefully generated pseudo-ground truth optical-flow map produced using a DEQ-Flow-based approach.

A notable design choice is that the endpoint frames used for VFI and OF are kept completely sharp. The paper explicitly argues that this makes interpolation more difficult because there are fewer blur-trajectory cues than in settings where endpoint frames are themselves motion-blurred. For evaluation, the benchmark uses PSNR and SSIM for deblurring, PSNR and SSIM for VFI, and EPE and F1 for OF. At the same time, the paper states that the OF labels are preliminary and need further refinement, so OF results are presented as proof-of-concept rather than definitive evaluation (Ciubotariu et al., 8 Sep 2025).

Sequence annotations include motion-pattern metadata such as subject motion, camera motion, translations on $O_x, O_y, O_z$, yaw, pitch, roll, and broader scene or acquisition diversity. This annotation structure is consistent with the benchmark’s role as both a restoration dataset and a controlled stress protocol over motion amplitude.

5. Benchmark structure and empirical behavior under increasing motion

The comparison table reports Size = 83,250 for VAR-MIORe, together with Max OF = 1932, Final FPS = [4, 1000], and Offset = [1, 249] ms (Ciubotariu et al., 8 Sep 2025). The benchmarked protocol visible in the experiments uses 333 input entries per bucket over the seven motion buckets $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$0, so the benchmarked subset visible there is 2331 entries. MIORe, by comparison, is reported as Size = 52,218, about 5% of the original sharp-frame pool.

The quantitative experiments show that existing methods degrade substantially as motion increases. For deblurring on VAR-MIORe, the paper reports PSNR trends such as AdaRevD: 39.61 $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$1 21.39, LoFormer: 41.41 $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$2 21.42, FFTformer: 35.33 $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$3 20.91, UFPNet: 37.83 $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$4 21.15, and NAFNet: 25.69 $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$5 20.03 across the seven blur buckets. The Input baseline goes from inf at 1 frame, then 34.32 $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$6 21.52. The paper highlights that at extreme blur, learned methods approach or even underperform the blurred input baseline in PSNR, indicating severe failure to recover information.

For VFI on VAR-MIORe, the same monotonic degradation appears. Reported PSNR trends include VFIMamba: 67.79 $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$7 18.91, SGM-VFI: 56.80 $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$8 18.71, PerVFI: 46.59 $$\|\mathbf{f}(\mathbf{x})\|_2 = \sqrt{u(\mathbf{x})^2 + v(\mathbf{x})^2},$$9 18.63, EMA-VFI: 81.16 $$\mu_f = \frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\mathbf{f}(\mathbf{x})\|_2, \qquad f_{\max} = \max_{\mathbf{x}\in\Omega}\|\mathbf{f}(\mathbf{x})\|_2.$$0 18.91, and BiFormer: 43.60 $$\mu_f = \frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\mathbf{f}(\mathbf{x})\|_2, \qquad f_{\max} = \max_{\mathbf{x}\in\Omega}\|\mathbf{f}(\mathbf{x})\|_2.$$1 18.90. The AVG(L,R) baseline goes from inf at 1-frame equivalent, then 30.41 $$\mu_f = \frac{1}{|\Omega|}\sum_{\mathbf{x}\in\Omega}\|\mathbf{f}(\mathbf{x})\|_2, \qquad f_{\max} = \max_{\mathbf{x}\in\Omega}\|\mathbf{f}(\mathbf{x})\|_2.$$2 18.22. The paper’s interpretation is that interpolation quality also collapses toward baseline quality as displacement grows.

The benchmark also provides split-wise analysis on MIORe itself. For deblurring, AdaRevD, LoFormer, and UFPNet are described as relatively balanced across MIORe splits; FFTformer shows a noticeable drop in the Mild-Medium group but appears to generalize better to fast motion, while NAFNet is more stable in slow scenes but weaker under extreme motion. For OF, the results are explicitly called preliminary; among reported MIORe numbers, MemFlow has the best total EPE = 2.99. No dedicated VAR-MIORe optical-flow quantitative table is shown in the provided text (Ciubotariu et al., 8 Sep 2025).

6. Limitations, significance, and research role

The paper states several limitations directly. The OF labels are pseudo-ground truth and require refinement. Although the corpus is captured in the wild, the final 333 sequences are selected, so the scene distribution is not guaranteed to be statistically representative of all real-world motion. The Chronos camera’s internal RAM constrains continuous capture to 5516 frames. Motion control is implemented indirectly through frame averaging and exposure offset over real captures rather than exact physical motion instrumentation. Finally, because motion blur coexists with defocus, exposure issues, lens artifacts, and adverse weather, failure cases cannot always be attributed purely to motion magnitude (Ciubotariu et al., 8 Sep 2025).

These caveats are integral to the benchmark’s meaning rather than peripheral to it. VAR-MIORe is presented as a real-world 1000 FPS, FHD, multi-task benchmark with explicit and controllable motion magnitude variation from minimal to extreme. Its importance lies in turning motion amplitude itself into a benchmark axis and in showing that methods that appear strong on conventional datasets often degrade sharply, or collapse toward baseline quality, when motion magnitude becomes large. The paper therefore frames VAR-MIORe not only as another dataset, but as a stress-testing protocol for robustness to motion scale, with direct implications for deblurring, VFI, and related restoration research.