{"id":"b74ae7b6-07f4-4426-bd6a-cb58df1d2e45","arxiv_id":"1909.00475","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors introduce visual deprojection and a conditional VAE that samples plausible images or videos consistent with a spatial or temporal projection.","lead":"Visual deprojection reconstructs images or videos that were collapsed into a lower-dimensional projection, such as a motion-blurred photo or a 1D edge-camera signal. This paper presents a probabilistic neural-network method that outputs multiple plausible originals for a given projection, which could aid seeing around corners and deblurring.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time samples are not guaranteed to be consistent with the input projection, yet consistency is the central promise of the method.","rationale":"The paper proposes a CVAE-style model for deprojection. Its novelty and potential usefulness are plausible, and the experimental setup is reasonable for a first exploration. The most load-bearing concern is not the choice of linear averaging per se, but that the objective never enforces projection consistency, even though the abstract and introduction repeatedly promise outputs 'consistent with a given input projection.' This is an internal gap between the stated claim and the loss function. It is concrete, testable, and directly affects whether the method fulfills its central promise. The reader's weakest assumption focuses on robustness to unknown/noisy projections, which is related but distinct. I therefore partially agree with the reader. The existing CONDITIONAL verdict is appropriate: the paper should not be fully accepted until either the consistency of generated samples is directly quantified or the objective is modified to include an explicit projection-consistency term. No harsher verdict is warranted because the method may still work in practice, and the missing evidence is readily obtainable.","tokens_in":10095,"tokens_out":2646,"duration_ms":29822,"concrete_test":"On the held-out FacePlace and Moving MNIST test sets, compute for each test projection x_i the projection residual of generated samples: r_i = min over sampled z of ||f_omega(g_theta(x_i, z)) - x_i||_2 / ||x_i||_2, and report the full distribution (median, 95th percentile) alongside DET and LMMSE. Separately, retrain with an explicit projection-consistency term lambda * ||f_omega(g_theta(x,z)) - x||^2 added to Eq. 6 and rerun the PSNR curves. If the median generated-sample projection PSNR is not above a pre-registered threshold (e.g., 30 dB) for the original model, or if adding the consistency term materially changes the samples, then the claim that test-time samples are consistent with the input projection is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that sampling from the inference network yields plausible candidates that are 'consistent with a given input projection.' The model, however, never explicitly enforces this consistency. In the generative model (Eq. 1), p_theta(y|x,z) is a Gaussian centered at g_theta(x,z) with no dependence on the known projection operator f_omega. The training loss (Eq. 6) contains only a KL term and a pixelwise reconstruction term ||g_theta(x, z_hat) - y||^2; there is no term such as ||f_omega(g_theta(x,z)) - x||^2. Since the true posterior p(y|x) is supported on the set {y : f_omega(y) = x}, a factorized Gaussian around an unconstrained decoder can put mass far from this set. The reported projection PSNR curves (Figs. 5 and 8) suggest that this is not merely a theoretical worry: DET and LMMSE often achieve higher projection PSNR than the proposed VDP method, and the paper does not report the fraction of samples that meet any specified projection-consistency tolerance. Without such a guarantee or an explicit constraint, the headline claim that outputs are 'consistent with the given projection' is not established by the objective or the evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces 'visual deprojection,' the task of reconstructing a higher-dimensional signal (image or video) from a lower-dimensional observation obtained by averaging along one dimension. The authors formulate a conditional variational autoencoder (CVAE) in which a latent variable z captures ambiguity in the mapping from projection x to signal y, with a learned conditional prior p_phi(z|x), a variational posterior q_psi(z|y), and a deprojection network g_theta(x,z). Training minimizes the CVAE ELBO with a reconstruction term and a KL regularization term, optionally augmented with a perceptual loss. Experiments cover spatial deprojection of face images (FacePlace), spatial deprojection of walking videos, and temporal deprojection of Moving MNIST videos from motion-blurred images. The authors report PSNR curves showing that sampling multiple candidates improves maximum signal PSNR relative to deterministic baselines, and they present qualitative results for the walking-video and Moving MNIST tasks. The central claim is that sampling from the inference network at test time yields plausible candidates consistent with the input projection.","tokens_in":10334,"tokens_out":5080,"duration_ms":49824,"significance":"If the claims are substantiated, the paper provides a unified probabilistic framework for an interesting and understudied task, with potential applications to corner cameras, motion blur, and other collapsed-dimension problems. The CVAE derivation is standard and correctly applied, and the paper is clearly written. The experimental design includes multiple datasets and several baselines (k-NN, deterministic network, LMMSE), which is a strength. However, the paper's most distinctive claim—that samples are 'consistent with a given input projection'—is not enforced by the objective and is only partially supported by the reported projection-PSNR curves. The lack of comparison to the closest prior work on recovering video from a single motion-blurred image also weakens the novelty claim. These issues are fixable but require additional analysis and experiments.","major_comments":[{"comment":"The training loss in Eq. (6) contains only the KL divergence and the reconstruction term ||g_theta(x,z_hat) - y||^2; there is no term such as ||f_omega(g_theta(x,z_hat)) - x||^2 that enforces consistency between generated samples and the observed projection. Because p_theta(y|x,z) in Eq. (1) is a Gaussian centered at g_theta(x,z) with no dependence on the known projection operator, samples from the model are not guaranteed to lie near the fiber {y : f_omega(y) = x}. The paper's abstract and Sec. 1 claim that sampling yields candidates 'consistent with a given input projection,' but this property is not established by the objective. The reported projection-PSNR curves in Figs. 5 and 8 indicate that DET and LMMSE often achieve higher projection PSNR than the proposed method, suggesting that the consistency issue is empirically relevant. The authors should report a direct projection-consistency metric (e.g., fraction of samples within a tolerance of the input projection) and either add an explicit consistency term to the loss or discuss why the learned model is expected to satisfy the constraint approximately.","section":"Sec. 4.3 and Sec. 4.4"},{"comment":"The temporal deprojection experiments (Moving MNIST) are not compared against the closest prior work on recovering video from a single motion-blurred image, namely Jin et al. [18] and Purohit et al. [30], both of which are cited in the related work. Without such a comparison, the claim that this is a 'first general method' and that it outperforms existing recovery approaches is not fully supported. In addition, the walking-video experiments in Sec. 4.3 are entirely qualitative; no quantitative metrics (e.g., PSNR, SSIM, or perceptual similarity on held-out subjects) are reported. The authors should either add quantitative evaluation for the walking videos or state explicitly that these results are preliminary qualitative demonstrations.","section":"Sec. 4.2"},{"comment":"The primary signal-reconstruction metric is the maximum PSNR over k samples. This metric rewards a method for producing a diverse or broad set of candidates and does not assess whether the distribution of samples is calibrated (e.g., whether the variability across samples matches the true posterior variability). Since the paper's contribution is explicitly probabilistic, the authors should report additional metrics that capture distributional quality, such as expected PSNR over samples, diversity measures, or a calibration curve comparing sample variability to ground-truth variability. Without such metrics, the probabilistic advantage over deterministic baselines is only partially demonstrated.","section":"Sec. 4.2"}],"minor_comments":[{"comment":"Equation (5) writes the log-likelihood as ||g_theta(x,z_hat) - y||^2 / (2 sigma_y^2) + const., but Eq. (6) drops the 1/(2 sigma_y^2) factor without explaining how sigma_y^2 is set. Please state the value or how it is absorbed into beta.","section":"Sec. 4.1"},{"comment":"The hyperparameter beta is selected so that the KL term lies in [5,15] on validation, and the latent dimension is fixed to 10. No sensitivity analysis is provided. A brief discussion of how these choices affect sample diversity versus reconstruction fidelity would improve reproducibility.","section":"Sec. 3.2"},{"comment":"The caption of Fig. 2 mentions 'N is the number of test examples in the dataset,' but N does not appear in the figure. Please clarify the notation or remove the reference.","section":"Fig. 2"},{"comment":"Several references lack publication venues or years (e.g., [17], [42], [45]). Please complete the bibliographic details.","section":"References"},{"comment":"The walking dataset description says 35 videos of 30 subjects, with 6 subjects held out, but does not specify the train/validation split used for hyperparameter selection. Please clarify.","section":"Sec. 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable CVAE application to an interesting problem, but the 'consistency' claim needs to be substantiated with a projection-consistency metric or an explicit constraint. The absence of comparisons with [18] and [30] on the motion-blur task is a notable omission, especially given the related-work section cites them. The walking-video section is qualitative only, which limits the strength of the spatial-deprojection claim. These issues are addressable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should look at this one if you care about inverse problems where a whole dimension is collapsed. The task definition alone is a real contribution: visual deprojection, recovering an image or video from a projection along space or time, is framed cleanly and the paper argues well that it is distinct from super-resolution, deblurring, and compressed sensing. The method is a conditional VAE with a latent variable, and the derivation of the ELBO is standard but correct. The authors also deserve credit for testing on three diverse settings (faces, gait, Moving MNIST) and for including non-trivial baselines like a nearest-neighbor selector and a linear MMSE estimator on the face task.\n\nNow the soft spots. The stress-test note is on target. The generative model p(y|x,z) has no dependence on the known projection operator f, and the training loss has no term like ||f(g(x,z)) - x||^2. So there is no mechanism that forces samples to be consistent with the input projection. The paper's own projection-PSNR plots show this is not a purely theoretical worry: DET and LMMSE often beat the proposed method on projection consistency, and the paper never reports what fraction of samples meet any given tolerance. This undercuts the headline claim that samples are \"consistent with the given input projection.\" It is fixable—add a consistency term, or at least report the metric—but it is a real gap.\n\nAlso, the paper cites two prior motion-blur-to-video methods (Jin et al. 2018, Purohit et al. 2019) but does not compare against them on the temporal deprojection task. That is a significant omission for a paper claiming a first general method. The walking-video experiments are qualitative only, and there are no error bars anywhere. No code or data is released, which makes the in-house gait dataset hard to build on.\n\nNone of these are fatal. The core idea is sound, and the framework is a plausible first solution to a new problem. The paper is worth a serious referee slot, but I would ask for a major revision before acceptance.\n\nFor you: if you work on corner cameras or motion deblurring, this is worth a look and worth citing for the task framing. I would bring it to a reading group—the consistency issue is a good case study in how conditional generative models can silently ignore known constraints.","headline":"The paper defines a genuinely new task and a sensible CVAE framework for it, but the central claim that samples are consistent with the input projection is neither enforced by the loss nor properly evaluated.","tokens_in":10869,"tokens_out":1870,"would_cite":true,"duration_ms":21941,"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":"The paper introduces visual deprojection and claims that a conditional variational autoencoder can recover the distribution of plausible originals behind a single collapsed-dimension projection.","keywords":["visual deprojection","motion blur inversion","conditional variational autoencoder","probabilistic image synthesis","video synthesis","corner cameras","long-exposure photography","projection inversion"],"falsifier":"Run the trained model on projections corrupted by sensor noise or an unknown blur kernel: if the sampled originals' reprojections consistently fail to match the observed projection, or if across many test cases the ground-truth original is rarely among the samples, the learned distribution does not cover the true posterior.","tokens_in":9911,"feed_emoji":"🎞️","tokens_out":5460,"duration_ms":49980,"temperature":0.7,"pith_summary":"The paper introduces visual deprojection: recovering an image or video after one of its dimensions has been averaged away, as in a motion-blurred photo or a corner camera's 1D feed. The task is ill-posed, so the paper's central claim is that a conditional variational autoencoder can model the full distribution of original signals consistent with a projection, not just one estimate. If the claim holds, a single probabilistic framework can tackle spatial collapses (faces, gait videos) and temporal collapses (video from long-exposure images) alike, and sampling at test time yields diverse plausible reconstructions. The paper evaluates this on face images, walking videos, and moving digits, showing that multiple samples improve the best reconstruction while preserving agreement with the input projection.","feed_headline":"One model un-blurs a collapsed dimension into many plausible originals","feed_subtitle":"Sampling from a learned conditional prior reconstructs faces, gait, and video from motion-blurred images.","key_machinery":"The load-bearing object is a conditional variational autoencoder combined with a known linear averaging projection operator, $x = \\sum_{k=1}^{d_p} \\omega_k y_{p=k}$. A strided 3D convolutional encoder models the variational posterior $q_\\psi(z|y)$, a 2D convolutional encoder models the conditional prior $p_\\phi(z|x)$, and a UNet-style decoder $g_\\theta(x,z)$ expands the projection along the collapsed dimension and refines the volume with 3D convolutions. The reparameterized ELBO loss ties test-time sampling to reconstruction accuracy and prior coverage, making the ambiguity of deprojection an explicit part of the model.","core_discovery":"Visual deprojection is stated as the task of estimating $p(y|x)$ for signals $y$ collapsed along a dimension by a known projection $x = f_\\omega(y)$; the paper proposes a conditional variational autoencoder in which a latent variable $z$ captures variability of $y$ not explainable by $x$, a prior $p_\\phi(z|x)$ maps projections to latent distributions, and a decoder $g_\\theta(x,z)$ synthesizes originals. Training maximizes an evidence lower bound whose loss is a $\\beta$ KL divergence between posterior and conditional prior plus reconstruction error, with an optional perceptual term for faces. At test time, sampling $z$ from the learned conditional prior and decoding produces multiple plausible originals rather than a single deterministic answer. The paper reports that this sampling strategy recovers face appearances from 1D projections, gait videos from vertical spacetime projections, and digit videos from heavily motion-blurred images, with maximum-signal PSNR improving as more samples are drawn.","pith_inferences":["A natural extension the paper leaves implicit is learning the projection operator itself, so the same ELBO objective could jointly infer blur kernels and scene distributions instead of assuming a fixed average.","The method's dependence on training data suggests a testable extension: train on paired high-speed and long-exposure real footage and measure how sample diversity and projection consistency change with dataset breadth.","The same conditional prior structure could apply to projectional radiography, where a 3D volume is collapsed into a 2D radiograph, though medical priors and measurement noise would need their own treatment."],"forward_implications":["One model can handle both spatial and temporal deprojection; any dimension collapsed by a linear average becomes a candidate target.","Sampling at test time turns a single ambiguous projection into a distribution over outcomes, so downstream tasks can reason about multiple plausible scenes rather than one estimate.","Recovering video from a motion-blurred image is feasible when training data covers the relevant dynamics, including the arrow-of-time ambiguity.","Drawing more samples improves the best reconstruction achievable from the model, while deterministic baselines saturate after one output."],"supporting_citations":[{"why":"Supplies the variational autoencoder objective and reparameterization trick used to derive the ELBO training loss.","marker":"[23]"},{"why":"Defines the conditional variational autoencoder formulation that conditions the latent prior on the input projection.","marker":"[36]"},{"why":"Establishes the corner-camera setting whose 1D spatial projections motivate 2D-to-3D deprojection.","marker":"[4]"},{"why":"A deterministic baseline that recovers a video sequence from one motion-blurred image, which the paper contrasts with its probabilistic approach.","marker":"[18]"},{"why":"Another deterministic motion-blur inversion baseline that the temporal deprojection experiments relate to.","marker":"[30]"},{"why":"Provides the Moving MNIST dataset used to construct motion-blurred projections for temporal deprojection evaluation.","marker":"[37]"},{"why":"Motivates the arrow-of-time ambiguity in temporal projections that the latent variable is designed to capture.","marker":"[43]"},{"why":"Provides the FacePlace face dataset used for 1D-to-2D spatial deprojection evaluation.","marker":"[31]"}],"fun_headline_variants":["Sampling a learned prior recovers many plausible originals from a projection","Deprojection: one input, many possible originals via conditional sampling","From blurry or 1D views, AI samples plausible reconstructions","One model, many originals: sampling handles collapsed-dimension vision tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the projection is a known, noiseless averaging of pixels along one dimension, identical at training and test time.","fun_headline_variants_meta":{"raw":{"variants":["Sampling a learned prior recovers many plausible originals from a projection","Deprojection: one input, many possible originals via conditional sampling","From blurry or 1D views, AI samples plausible reconstructions","One model, many originals: sampling handles collapsed-dimension vision tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000726,"raw_usage":{"total_tokens":3249,"prompt_tokens":933,"completion_tokens":2316,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":2237}},"tokens_in":549,"tokens_out":2316,"duration_ms":14982,"temperature":1.0,"reasoning_tokens":2237,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:51:35.709839+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on projections corrupted by sensor noise or an unknown blur kernel: if the sampled originals' reprojections consistently fail to match the observed projection, or if across many test cases the ground-truth original is rarely among the samples, the learned distribution does not cover the true posterior.","supporting_citations":[{"cited_title":"Learning structured output representation using deep conditional gen- erative models","cited_arxiv_id":null,"evidence_quote":"Defines the conditional variational autoencoder formulation that conditions the latent prior on the input projection."},{"cited_title":"Turning corners into cameras: Princi- ples and methods","cited_arxiv_id":null,"evidence_quote":"Establishes the corner-camera setting whose 1D spatial projections motivate 2D-to-3D deprojection."},{"cited_title":"Learning to Extract a Video Sequence from a Single Motion-Blurred Image","cited_arxiv_id":"1804.04065","evidence_quote":"A deterministic baseline that recovers a video sequence from one motion-blurred image, which the paper contrasts with its probabilistic approach."},{"cited_title":"Bring- ing alive blurred moments","cited_arxiv_id":null,"evidence_quote":"Another deterministic motion-blur inversion baseline that the temporal deprojection experiments relate to."},{"cited_title":"Unsupervised learning of video representations using lstms","cited_arxiv_id":null,"evidence_quote":"Provides the Moving MNIST dataset used to construct motion-blurred projections for temporal deprojection evaluation."},{"cited_title":"Learning and using the arrow of time","cited_arxiv_id":null,"evidence_quote":"Motivates the arrow-of-time ambiguity in temporal projections that the latent variable is designed to capture."},{"cited_title":"Recogniz- ing disguised faces","cited_arxiv_id":null,"evidence_quote":"Provides the FacePlace face dataset used for 1D-to-2D spatial deprojection evaluation."}],"review_version":1}