{"id":"82b82b9c-73a7-4deb-b0f6-72d635f55957","arxiv_id":"1908.01940","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Combining compressed sensing of sparse tracked trajectories with polynomial-expansion optical flow restores wavy-water-distorted video more accurately and much faster than earlier state-of-the-art methods.","lead":"This paper restores videos of underwater scenes distorted by waves by tracking a few salient points, estimating the full motion field with compressed sensing, and applying a simple polynomial optical flow step. The combined method reports lower error and sharper structure than three previous state-of-the-art approaches on synthetic and real videos.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"COT bias is the load-bearing risk: if the water motion has a nonzero temporal mean (current, drift, or asymmetric waves), Eq. 1 receives systematically biased measurements and the restoration claims rest on an untested assumption.","rationale":"The reader's weakest_assumption identifies exactly this COT bias, and I agree it is the most serious threat to the central claim. The CS pipeline's entire input is displacement trajectories computed relative to a time-averaged center; if that center is biased, every measurement in Eq. 1 is systematically wrong, not merely noisy. The paper's own validation does not probe this: synthetic waves are zero-mean sinusoids, and the real tank experiments use mechanical paddles that likely generate near-zero-mean standing waves. The filtering of unstable trajectories does not remove a constant bias shared by all frames. The concrete test I propose directly varies the temporal mean of the displacement while keeping all other components unchanged, isolating the COT assumption. If the restoration metrics stay robust up to several pixels of bias, the concern is not fatal; if they degrade sharply, the method's claimed generality is not supported. The reader's CONDITIONAL verdict remains correct, since this is an addressable empirical question rather than a proven flaw in the reported experiments.","tokens_in":13006,"tokens_out":8858,"duration_ms":104522,"concrete_test":"Create a synthetic sequence with the same refraction model as Sec. 3.1 but add a controlled bias to the water surface: for example, add a steady current term c*t to the surface height (or equivalently add a constant offset to every frame's displacement field), with the total bias |c*T| ranging from 0 to 5 pixels. Run the released code for CS and CS+PEOF on these sequences. Plot restored-image RMSE and SSIM (relative to the known ground truth I0) as functions of the bias magnitude. If RMSE degrades by more than ~0.05 or SSIM drops by more than ~0.02 for a 2-pixel bias, the COT assumption is not robust. Also run a zero-mean-wave control to confirm the pipeline reproduces the paper's reported performance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is the center-of-trajectory (COT) approximation in Sec. 2.3.2. The displacement measurements feeding Eq. 1 are d_i(t) = p_i(t) - COT_i, with COT_i = mean_t p_i(t). For the CS model to be correct, COT_i must equal the true undistorted coordinate p_i^0; otherwise every measurement is d_i^true(t) - b_i, where b_i = COT_i - p_i^0 is the temporal mean displacement at that point. The paper justifies this by 'local symmetry of water motion' but provides no stress test. A net current, Stokes drift, or an asymmetric wave train produces a nonzero, possibly spatially varying b_i. This is not caught by the paper's trajectory filtering rule (COT over first vs. last T/2 differing by >3 px), because a constant offset adds equally to both halves. The biased measurements propagate through LASSO (Eq. 2) into a biased MVF, and the warped frames produce a mean image that is globally shifted or geometrically distorted relative to I0. All restoration metrics in Table 1 would then be compared against a misregistered reference, so the central comparison is confounded under such conditions. This is the fulcrum of the pipeline: if COT breaks, the CS part and hence CS+PEOF have no valid input. The paper's real data come from a mechanically-paddled tank, where waves are likely near-zero-mean, so this failure mode is untested outside that narrow regime.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses restoration of a static scene imaged through a wavy water surface, where each video frame is a non-rigidly distorted view of the same underlying image. The authors propose a compressed sensing (CS) pipeline: salient points are tracked across frames, their displacement trajectories are used as sparse space-time measurements of the dense motion vector field (MVF), and the MVF is recovered by solving a LASSO problem under the assumption that the MVF is sparse in the 3D DFT basis. The restored image is then obtained by warping each frame by the estimated MVF and taking the mean. The paper also evaluates a simpler optical flow method based on local polynomial expansion (PEOF), and a two-stage CS+PEOF variant. Experiments on synthetic videos, a mechanically-paddled real tank dataset, and the Real2 dataset of Tian and Narasimhan report numerical improvements over SBR-RPCA, LWB, and a deep-learning baseline in terms of RMSE, SSIM, and NMI, with substantially lower computation time than SBR-RPCA and LWB.","tokens_in":13305,"tokens_out":5297,"duration_ms":61105,"significance":"If the reported results hold, the CS-based MVF interpolation is a genuinely new use of compressed sensing for refractive distortion correction, and the finding that a simple PEOF optical flow method is competitive with or better than dedicated state-of-the-art restoration methods is practically valuable. The paper is refreshingly clear about its assumptions, provides synthetic and real data with ground truth, and makes code and datasets available, which strengthens reproducibility. The central claims, however, depend on two assumptions that are not independently verified: the zero-mean local symmetry of water motion that justifies the center-of-trajectory approximation, and the empirical sparsity of real MVFs in the DFT basis. The current evidence for the latter is circular, and the former is only tested in a wave-tank regime where waves are generated by mechanical paddles and are likely near-zero-mean. The CS-theoretic justification also contains a coherence error.","major_comments":[{"comment":"The empirical verification of Fourier sparsity is circular. The MVF whose DFT coefficients are plotted in Fig. 1 is the output of the CS/LASSO estimator in Eq. (2), which explicitly minimizes the ℓ1 norm of the DFT coefficients and therefore enforces sparsity in the DFT basis. Showing that this estimate has a concentrated spectrum is a tautology, not evidence that real MVFs are sparse in the DFT basis. The authors do note that 'the actual MVF can only be estimated,' but they then use the CS estimate itself to support the modeling assumption. Please verify sparsity using an independently estimated MVF (for example, from PEOF/optical flow, or from the known synthetic MVF in the Synthetic dataset), or explicitly reword the section to state that Fig. 1 illustrates the estimator's output rather than validating the premise on which the estimator is built.","section":"Sec. 3.3, 'Verification of Fourier Sparsity' and Fig. 1"},{"comment":"The COT approximation x_i0 ≈ mean_t x_it is load-bearing for the entire CS stage, because every displacement in Eq. (1) is computed relative to the COT. The justification by 'local symmetry of water motion' fails when there is a net current, Stokes drift, or an asymmetric wave train: each trajectory then has a bias b_i = COT_i - p_i^0, and all measurements e in Eq. (1) are systematically shifted by b_i. The trajectory-filtering rule in Sec. 2.3.1 (COT over the first vs. last T/2 frames differing by more than 3 pixels) cannot detect this bias, because a constant offset adds equally to both halves. All real data come from a mechanically-paddled tank, where waves are plausibly near-zero-mean, so the failure mode is untested. Please add a synthetic experiment with a nonzero-mean velocity component (a constant drift or a biased wave) and report the restoration metrics and the resulting motion reduction; if the method degrades, state the operating domain clearly, and if it does not, explain why the bias is not harmful in CS reconstruction.","section":"Sec. 2.3.2, center-of-trajectory (COT) approximation, with Sec. 2.3.1 filter"},{"comment":"The claim that the coherence between the row-subsampled identity sampling matrix Φ and the DFT basis F 'reaches its lower bound of 1' is numerically incorrect as stated. With the definition given in the paper, |Φ_i F_j| = 1/√(NxNyT) for every pair of rows/columns, so the defined coherence is approximately 1/√(NxNyT), not 1; under the standard sensing-matrix coherence used in CS theory, the value depends on the chosen row set and for a random partial DFT is typically O(√(log(NxNyT))), not exactly 1. The lower bound of coherence is also not 1 in this setting. Please correct the coherence calculation and the resulting sample-complexity sentence; otherwise the CS-theoretic justification for the number of tracked points is unsupported. This is a local error in the theory section, but it is part of the key novelty claim, so it should be fixed.","section":"Sec. 2.3.3, coherence argument after Eq. (1)"},{"comment":"The central empirical claim that CS, PEOF, and CS+PEOF outperform SBR-RPCA, LWB, and DL rests on single runs without error bars, repeated trials, or statistical tests. Several of the differences are small (e.g., Real1 Checker: CS+PEOF SSIM 0.892 vs. PEOF 0.890; Synthetic BrickWall: CS+PEOF RMSE 0.141 vs. PEOF 0.142), and all of the Synthetic videos are generated from 'randomly chosen parameters' with no indication of how many random realizations were used. Please report means and standard deviations over at least several random wave/tracking realizations for the Synthetic dataset, or otherwise characterize the run-to-run variability of the metrics. Without this, the superiority claims over the baselines are not yet supported at the level the paper asserts.","section":"Table 1 and Sec. 3.3"}],"minor_comments":[{"comment":"The LASSO objective in Eq. (2) uses an unsquared data-fidelity term ||e−ΦFθ||_2, whereas the standard LASSO and the YALL1 solver typically use a squared ℓ2 term. Please clarify the exact optimization problem solved by YALL1 in this paper, and ensure the objective function matches the solver's formulation.","section":"Sec. 2.3.3, Eq. (2)"},{"comment":"The description of the Synthetic dataset says the water surface is generated using a superposition of 2–6 sinusoidal waves with 'randomly chosen parameters,' but no details are given on the parameter ranges or the number of random trials. Adding this information would help reproducibility and would support the request for error bars in the major comments.","section":"Sec. 3.1 and Sec. 3.2"},{"comment":"The caption of Fig. 2 lists the datasets used but does not identify which curve corresponds to which dataset or parameter; the graph itself is not visible in the text. Please add a legend or descriptive labels, and state the axes clearly.","section":"Fig. 2"},{"comment":"The motion-reduction measure MR is defined as a median over tracked points, but the denominator ||d_i||_2 is the norm of the original displacement trajectory; for points with very small displacement this ratio can be unstable. Please state whether any thresholding or regularization was applied, and report the distribution of the per-point ratios rather than only the median.","section":"Sec. 3.3, 'Verification of Fourier Sparsity'"},{"comment":"There are several small typographical and formatting issues, such as 'Y ALL1' with a space in Sec. 3.2, inconsistent time formatting in Table 1 (e.g., '29s' vs. '0m 23s'), and the phrase 'nonrigid transport object' in reference [19]. These do not affect the technical content but should be cleaned up.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution with code and data provided, but the circular sparsity verification and the untested COT bias are real correctness concerns for the central claims. Both are fixable with additional experiments (an independent sparsity check and a drift/current stress test), and the coherence error is a localized but load-bearing theoretical mistake. I would support a major revision rather than rejection, since the proposed corrections are within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know about this paper: it's a genuine new idea -- using compressed sensing with a global 3D DFT sparsity prior to interpolate a dense non-rigid motion field from a few tracked trajectories -- and it ships code and data, with consistent wins over three published baselines on a 14-video test bed. The central claim holds up on their data; the biggest caveat is a load-bearing assumption that doesn't get stress-tested.\n\nWhat's actually new: the CS-based MVF interpolation is, as far as I know, not in the prior literature. The prior sparsity-based optical flow works use local patches and two-frame brightness constancy; this works globally in space-time and uses displacement trajectories as measurements. That's a real contribution. The PEOF result is also useful: a simple, fast optical flow method outperforms more expensive SOTA on this task, and the two-stage CS+PEOF is better than either alone. The experiments are extensive: synthetic and real videos, motion reduction numbers, parameter sensitivity, and compute time. The paper is clearly written.\n\nWhere the soft spots are:\n\n1. The COT assumption: Sec 2.3.2 equates the time-average of each trajectory with the undistorted coordinate, justified by 'local symmetry of water motion.' If there's a net current, Stokes drift, or an asymmetric wave train, every displacement measurement inherits a spatially varying bias, and the CS model is fed systematically wrong data. Their trajectory filtering rule (COT over first vs second half differing by >3px) won't catch a constant offset. Their real data comes from a mechanically-paddled tank, where waves are near-zero-mean, so this is untested. This is the paper's main limitation. It's not fatal, because zero-mean symmetric wave motion is a plausible and explicitly stated assumption, but it deserves a stress test (simulate drift, or find real data with current).\n\n2. The Fourier sparsity 'verification' in Sec 3.3 is circular: they show the MVF estimated by their CS method (which assumes DFT sparsity) has sparse DFT coefficients. That doesn't corroborate the model independently. They could run PEOF or any other dense flow method and check its DFT spectrum.\n\n3. No error bars on any metric. The gains are consistent across many videos, so this is minor, but repeated synthetic trials with random wave parameters would help.\n\n4. The comparison to [15] is done across papers, using different sequences (their 'Middle/Small/Tiny' vs [15]'s 'Large/Medium/Small'). That's not a valid head-to-head. Either run the code or drop the claim.\n\nBottom line: the paper is a solid contribution, worth a serious referee. The COT issue is a revision-level concern, not a reject-level one. If the authors address the stress test and the circular verification, I'd be happy to see it in a good venue.\n\nRecommendation: engage with it.","headline":"A genuinely new CS-based motion interpolation method with reproducible wins over baselines, but the center-of-trajectory assumption is a real, untested limitation.","tokens_in":13846,"tokens_out":3970,"would_cite":true,"duration_ms":80231,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that wavy-water video distortion can be undone by recovering the full motion field from a few tracked points using compressed sensing with a 3D Fourier sparsity basis.","keywords":["underwater image restoration","non-rigid distortion","compressed sensing","motion vector field","3D discrete Fourier transform","optical flow","polynomial expansion","trajectory tracking"],"falsifier":"Record a static target through wavy water with a known steady current or asymmetric wave train, track feature points, and compare each trajectory's mean against its known true position; if the mean offset exceeds a small threshold, the compressed-sensing reconstruction is fed biased measurements and the restored-image error should rise, whereas an algorithm that explicitly estimates drift would not degrade.","tokens_in":12793,"feed_emoji":"🌊","tokens_out":4366,"duration_ms":45299,"temperature":0.7,"pith_summary":"The paper tries to establish that a static scene filmed through wavy water can be restored by exploiting the spatio-temporal smoothness and periodicity of water waves. It tracks a small number of feature points, converts their trajectories into displacement measurements, and uses compressed sensing with a 3D discrete Fourier basis to estimate the dense motion field at every pixel. A simple optical-flow refinement based on local polynomial expansion is then applied, sometimes alone and sometimes after the compressed-sensing step. The authors claim these methods restore both synthetic and real underwater videos with lower error and better structural preservation than existing restoration techniques, while needing no training data, no distortion-free template, and no special hardware. A sympathetic reader would care because this offers a physically grounded, data-free route to a practical vision problem that interferes with monitoring, recognition, and tracking in underwater scenes.","feed_headline":"Wavy-water distortion solved from a few tracked points","feed_subtitle":"Treating the motion field as a sparse 3D Fourier signal restores underwater video faster and with less distortion than prior methods.","key_machinery":"The central object is the complex-valued motion vector field $d = d_x + i d_y$, a 3D array over image space and time. Its key property is sparsity in the 3D DFT basis, justified by the spatial smoothness and temporal periodicity of water waves. The measurement model is $e = \\Phi F \\theta + \\eta$, where $\\Phi$ is a row-subsampled identity selecting tracked points and $F$ is the 3D DFT matrix; because $\\Phi$ and $F$ are maximally incoherent, a LASSO solver can reconstruct the field from very few tracks. The second ingredient is PEOF, which fits local second-degree polynomials to image patches and solves for the displacement between each frame and the mean image, iterating over a pyramid for large motions.","core_discovery":"The central claim is that the full 3D motion vector field generated by a wavy water surface is compressible in the 3D discrete Fourier basis, so it can be recovered from a small set of tracked point trajectories by solving an L1-regularized least-squares problem, the LASSO. The recovered field is then used to warp every video frame back toward a common undistorted coordinate system. The paper further claims that applying polynomial-expansion optical flow (PEOF) after the compressed-sensing step preserves straight lines, text, and grid structure better than the compressed-sensing step alone, and that both stages outperform the compared existing methods on synthetic and real underwater videos in terms of RMSE, SSIM, and normalized mutual information. The two-stage CS+PEOF pipeline is presented as the best overall restoration approach, at computational cost far below the spline-plus-RPCA baseline.","pith_inferences":["If the zero-mean water-motion assumption fails because of a steady current or wave drift, the center-of-trajectory estimate becomes biased; a natural extension would be to estimate and subtract a global or slowly varying drift term before compressed-sensing reconstruction.","The same 3D-DFT sparsity idea could be tested on other temporally periodic non-rigid distortions, such as short-exposure atmospheric turbulence, where similar band-limited motion fields arise.","The success of this parameter-free pipeline suggests that deep single-frame restoration methods lose information by ignoring temporal redundancy; training them on multiple frames, or combining them with a compressed-sensing motion-reduction front end, could close the performance gap.","Because the row-subsampled identity and the DFT are maximally incoherent, the reconstruction guarantees are near-optimal; replacing the DFT with another basis would likely require more tracked points for the same accuracy."],"forward_implications":["Dense non-rigid motion estimation for wavy water needs only a sparse set of tracked points: the paper reports reconstructions from fewer than 3% of pixel trajectories, with performance saturating as more tracks are added.","The compressed-sensing stage alone removes more than 90% of the motion in most test videos, indicating that the 3D-DFT sparsity model captures the physical distortion well.","A simple local-polynomial optical-flow stage, used alone, already beats the compared baselines on most numerical metrics, so elaborate models are not required for this task.","The two-stage CS+PEOF combination gives the best structural preservation of lines, text, and grids, as well as the best numerical quality in most comparisons.","The method needs no training data, no distortion-free template, and no special illumination or multiple viewpoints; only a short video on the order of 100 frames is sufficient."],"supporting_citations":[{"why":"Supplies the refraction-based image formation model that relates water-surface derivatives to per-pixel displacements.","marker":"[19]"},{"why":"Provides the compressed-sensing recovery guarantees that justify reconstructing a sparse signal from few incoherent measurements.","marker":"[8]"},{"why":"The YALL1 LASSO solver used to estimate the sparse Fourier coefficients in Eq. 2.","marker":"[4]"},{"why":"Defines the polynomial-expansion optical flow (PEOF) method used as the second stage.","marker":"[13]"},{"why":"The spline-registration plus robust-PCA baseline that the paper compares against and outperforms.","marker":"[20]"},{"why":"Provides the learned water-bases baseline and the real text sequences used for evaluation.","marker":"[29]"},{"why":"The deep-learning single-frame restoration baseline that the paper argues ignores temporal redundancy.","marker":"[18]"},{"why":"The dense optical-flow interpolation method compared for motion estimation from sparse correspondences.","marker":"[22]"}],"fun_headline_variants":["Sparse tracking restores wavy-water video distortions","Wavy water corrected via compressed sensing and optical flow","Few tracked points undo wavy-water distortion via CS+flow","Compressive sensing plus polynomial flow cleans underwater video","Underwater video restored from sparse tracked motion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the time-average of each tracked point's positions equals its true undistorted position, which requires the water surface motion at every point to be locally symmetric and zero-mean over the captured frames.","fun_headline_variants_meta":{"raw":{"variants":["Sparse tracking restores wavy-water video distortions","Wavy water corrected via compressed sensing and optical flow","Few tracked points undo wavy-water distortion via CS+flow","Compressive sensing plus polynomial flow cleans underwater video","Underwater video restored from sparse tracked motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001648,"raw_usage":{"total_tokens":6557,"prompt_tokens":966,"completion_tokens":5591,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":5516}},"tokens_in":582,"tokens_out":5591,"duration_ms":37396,"temperature":1.0,"reasoning_tokens":5516,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:58:45.461763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a static target through wavy water with a known steady current or asymmetric wave train, track feature points, and compare each trajectory's mean against its known true position; if the mean offset exceeds a small threshold, the compressed-sensing reconstruction is fed biased measurements and the restored-image error should rise, whereas an algorithm that explicitly estimates drift would not degrade.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the refraction-based image formation model that relates water-surface derivatives to per-pixel displacements."},{"cited_title":"Candes, J","cited_arxiv_id":null,"evidence_quote":"Provides the compressed-sensing recovery guarantees that justify reconstructing a sparse signal from few incoherent measurements."},{"cited_title":"http://yall1","cited_arxiv_id":null,"evidence_quote":"The YALL1 LASSO solver used to estimate the sparse Fourier coefficients in Eq. 2."},{"cited_title":"Farneb ¨ack","cited_arxiv_id":null,"evidence_quote":"Defines the polynomial-expansion optical flow (PEOF) method used as the second stage."},{"cited_title":"Oreifej, G","cited_arxiv_id":null,"evidence_quote":"The spline-registration plus robust-PCA baseline that the paper compares against and outperforms."},{"cited_title":"Tian and S","cited_arxiv_id":null,"evidence_quote":"Provides the learned water-bases baseline and the real text sequences used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The deep-learning single-frame restoration baseline that the paper argues ignores temporal redundancy."},{"cited_title":"Revaud, P","cited_arxiv_id":null,"evidence_quote":"The dense optical-flow interpolation method compared for motion estimation from sparse correspondences."}],"review_version":1}