Revisiting Ground-Truth Synthesis from High-Speed Video: Exact Validity Conditions and an Audited Consumer Capture Corpus

Abdullah Al Shafi; Sumaiya Rahim Suma

cs.CVOct 4, 2026

Revisiting Ground-Truth Synthesis from High-Speed Video: Exact Validity Conditions and an Audited Consumer Capture Corpus

Authors: Abdullah Al Shafi, Sumaiya Rahim Suma

Organizations: Department of Computer Science and Engineering Khulna University of Engineering & Technology, Khulna, Bangladesh

Abstract

Motion-deblurring datasets are commonly synthesised by averaging $N$ consecutive high-frame-rate frames and labelling the result with the central frame. We show that this label is unbiased for every capture timing only when the window is odd and the frames' sample durations are equal. An even window shifts every label by a fixed fraction of the blur length, even under perfect timing. Unequal durations are subtler: their mean misalignment is zero, so dataset statistics cannot reveal them, yet when the offset cannot be read from the blur they convolve the supervision rather than adding noise to it. Auditing 51 smartphone clips recorded at a nominal 240 fps, we find 19 captured near 176 fps, a behaviour recorded only in the container's timing tables. In a controlled test the parity choice costs a fitted linear deblurring filter far more than these timing irregularities do, and interpolation labels, unlike blur labels, can be repaired with the true frame times. We provide a tool that checks both conditions without decoding, and will release the clips and their timing tables.

Figures & tables

Figure 1: The sharp label against the centroid of the blur it is paired with. (a) Odd window, equal durations: aligned. (b) Even window: a fixed offset in every pair. (c) Held intervals: an offset that varies and averages to zero.

Violation	Blur length	Cost (dB)	Learned shift
Even window, random direction	16–32 px	$0.42\pm 0.05$	none: offset becomes blur
Even window, random direction	32–48 px	$0.73\pm 0.07$	none
Even window, direction fixed	16–32 px	$0.85\pm 0.03$	1.08 of the 1.72-px offset
Real hold patterns	8–48 px	$0.04$ to $0.05$	none

Table 1: What each violation costs a fitted linear deblurring filter: PSNR loss of the conventional label against the frame at the true centroid, mean and standard deviation over five seeds, with the net shift the filter learns.

Figure 2: The corpus at a glance. (a) Effective capture rate per clip; the dashed line is the nominal 240 fps, which uniform clips meet and affected clips miss at 174–179 fps. (b) Share of intervals at each sample duration (log axis): one ten-tick slot for uniform clips; a second mode at twenty ticks, exactly twice nominal, for affected ones. The

\pm 1

tick spread is quantisation dither. (c) Held intervals, one mark each, in 320 consecutive mid-recording intervals of four affected and two uniform clips: clusters separated by clean runs, and none in uniform clips.

Figure 3: Why the timing is real. (a) Frames present in each clip against the number a uniform 240 fps stream of the same duration would hold: affected clips fall well below the identity line, so frames are missing. (b) Residual of cumulative timestamps from a linear fit, three clips per population on separate scales: affected residuals wander far outside the

\pm 0.75

tick band (shaded); uniform ones are the bounded sawtooth of dither.

	Uniform	Affected
Clips / duration / frames	32 / 26.8 min / 385,263	19 / 33.7 min / 356,656
Sample durations (ticks)	$\{9,10,11\}$	$\{9,10,11,19,20,21,30\}$
Share at one / two nominal slots	100.00% / —	64.16% / 35.82%
Effective rate	239.90 fps	174–179 fps
Linear-fit residual (ticks)	0.52–0.53	47.97–417.26

Table 2: The corpus separates under identical device, operating system and nominal settings. Residuals are the maximum absolute deviation of cumulative timestamps from a linear fit, in ticks; rounding alone stays near half a tick. Frame counts sum to the corpus total of 741,919.

Hypothesis	It predicts	The corpus shows	Verdict
Padded stream	repeated, near-empty frames	$0.00\%$ near-duplicates; no frame under $2\%$ of median size; post-hold frames no smaller (ratio $1.10$ )	excluded
Mislabelled durations	clock disagreement; motion ratio 1	clock agrees to 2 ms; motion across holds $1.72\times$ normal (IQR $1.59$ – $1.94$ , nine clips with flow above $0.05$ px at 480-px width, threshold set post hoc)	excluded
Drift within a clip	held fraction changes	$36.30\%$ to $35.91\%$ by decile (Wilcoxon $p=0.070$ )	no detectable drift
Thermal build-up	rises with duration, recovers	none over 15.4–301.9 s; $35.8\%$ after a 46-hour gap	not supported
Low light (Auto FPS)	affected clips darker	mean luminance $139.6$ (affected) vs $132.3$ (uniform) on a 0–255 scale (Mann–Whitney $p=0.108$ )	not supported
Visible to a builder	headers or frames differ	identical headers, metadata, codec; sharpness median $\approx 2500$ in both	not detectable

Table 3: Evidence and ruled-out explanations. Each row is a hypothesis that would produce an uneven timing table, what it predicts, and what the corpus shows. Session-level comparisons are descriptive: with two sessions, clip-level tests carry no session-level weight.

Figure 4: How held intervals are placed. (a) Gaps between consecutive held intervals, pooled over the affected population. Mass at one and two frames, a second mode beyond ten, and nothing between: the signature of clustered placement. (b) Autocorrelation of the held-interval indicator: median across all 19 affected clips, against the simultaneous

95\%

band of a permutation null pooled over clips (grey), the band the significance tests use. (c) Lengths of runs of consecutive held intervals.

Figure 5: The model against the audit. (a) Misalignment under perfect timing (Eq. 2 ): zero for odd

N

1/(2(N-1))

for even

N

. (b) Measured median and p95 misalignment in the affected population (solid) against P3 at

p=0.359

(dashed), treating the timing term as Gaussian (median

0.674\sigma

, p95

1.96\sigma

): the scale agrees, but the median rises with

N

where P3 predicts a fall, because holds are not independent. (c) Cumulative distribution of

\delta

N=7

: a point mass at zero for uniform clips; up to

\delta\approx 0.1

for affected ones.

Population	Stride	signed mean	sd	$\|{\cdot}\|$ p50	p75	p95
affected	4	$-2.0\times 10^{-6}$	0.0680	0.0385	0.0833	0.1364
affected	8	$-1.2\times 10^{-5}$	0.0856	0.0833	0.1087	0.1364
affected	16	$-1.0\times 10^{-5}$	0.0346	0.0322	0.0455	0.0581

Table 4: Interpolation label error

t_{\text{true}}-0.5

in the affected population (uniform: median zero, p95 at most

0.0031

). The signed mean vanishes; the p95 coincides at strides 4 and 8, which reach the same attainable ratio,

t=7/11

Explore similar work

Jun 24, 2025cs.CV

SloMoDeblur: A Large-Scale Smartphone Image Deblurring Dataset

Motion blur remains one of the most common and visually disruptive degradations in real-world smartphone imaging, yet existing deblurring benchmarks are often limited in scale, resolution, or domain relevance. This gap is especially pronounced for smartphones, where rolling shutter, small sensors, and ISP processing produce blur statistics that differ from GoPro/DSLR-based benchmarks. We introduce a large-scale smartphone-oriented deblurring dataset constructed from 240~fps slow-motion video. To approximate exposure-time radiance integration, we synthesize blur by temporally averaging a fixed window of

N=30

consecutive frames, which corresponds to an effective exposure of

T=1/8

~second, and we select the temporally centered frame as the sharp ground truth. The resulting benchmark contains 42,045 paired blur--sharp images at

1920\times1080

resolution spanning 843 distinct scenes, with a train/test split of 37,841/4,204 pairs. We benchmark multiple state-of-the-art deblurring models using PSNR and SSIM and observe consistent performance degradation relative to the baseline similarity between the input blurry images and ground truth, underscoring the realism and difficulty of the proposed data. We release the dataset and generation scripts via HuggingFace to facilitate the development and evaluation of robust, deployment-oriented deblurring methods.

Syed Mumtahin Mahmud, Mahdi Mohd Hossain Noki, Prothito Shovon Majumder +4

Cognitive Agents and Interaction Lab, Department of Computer Science and Engineering, University of Dhaka, Dhaka-1000, Bangladesh

Gaussian Deblurring Smartphone Photographs

Ask PDF

Oct 1, 2026cs.CV

PickMoment: Continuous-Time Single-Image-to-Video via Learning Deblurring and Blur-to-Video

Motion blur arises from the temporal integration of a continuous sharp signal over a finite exposure window, yet existing learning-based methods sidestep this physical model and predict only the sharp signal itself: most single-image deblurring methods recover a single frame at the exposure center, while blur-to-video methods predict a fixed set of frames. We introduce PickMoment, a continuous-time reformulation that directly learns the interval-mean blur over arbitrary sub-intervals of the exposure with a single deterministic model. Drawing an analogy to MeanFlow's average-velocity formulation, we train the model with three supervisions derived from the blur integral: an empirical reconstruction loss from available subframes, an additivity loss that enforces self-consistency across overlapping sub-intervals, and a sharp-frame loss anchored at the zero-interval limit. A single trained model unifies single-image deblurring, blur-to-video generation, and continuous-time pick-a-moment recovery as different queries to the same network, with no separate training for each task. Our PickMoment achieves state-of-the-art performance among generative-based deblurring methods on GoPro and HIDE while competitive against restoration-based methods on RealBlur, and the highest per-frame fidelity on GoPro-7 blur-to-video, all in a single forward pass without iterative sampling.

Junseong Shin, Hyeonsu Jo, Daehyun Kim +1

Department of Artificial Intelligence, Hanyang University · Department of Intelligence Convergence, Hanyang University · Department of Computer Science, Hanyang University

Gaussian Deblurring Motion Blur Video Restoration Single Image Dehazing

Ask PDF

May 7, 2026cs.CV

iPhoneBlur: A Difficulty-Stratified Benchmark for Consumer Device Motion Deblurring

Motion blur restoration on consumer mobile devices is typically evaluated using aggregate metrics that obscure performance variation across blur difficulty, masking model behavior under real deployment conditions. This work introduces iPhoneBlur, a difficulty-stratified benchmark of 7,400 image pairs synthesized from high-framerate iPhone 17 Pro videos captured in diverse real-world scenarios. Samples are partitioned into Easy, Medium, and Hard categories through PSNR-guided adaptive temporal windowing, with stratification validated by monotonic 2.2x increase in optical flow magnitude across tiers. Each sample includes comprehensive metadata enabling investigation of ISP-aware and difficulty-adaptive restoration strategies. Spectral analysis confirms synthesized blur exhibits high-frequency suppression patterns consistent with authentic motion degradation. Evaluation of six architectures reveals consistent 7-9 dB performance degradation from Easy to Hard subsets, a substantial gap entirely hidden by aggregate reporting. The benchmark further exposes a domain gap between professional and consumer cameras which targeted fine-tuning substantially recovers. By coupling difficulty stratification with deployment-critical metadata, iPhoneBlur enables systematic assessment of model reliability and failure modes for resource-constrained edge systems.

Abdullah Al Shafi, Kazi Saeed Alam

Department of Computer Science and Engineering Khulna University of Engineering & Technology, Bangladesh

Motion Blur Camera Resource-Constrained Edge Devices Motion

Ask PDF