{"id":"c91b527a-0734-4cee-b7e0-b41c84475e31","arxiv_id":"2608.13502","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"GS2CI recovers a static 3D Gaussian scene and its camera trajectory from one SCI snapshot by combining vision-foundation-model initialization with an opacity-guided densification strategy.","lead":"A new pipeline reconstructs 3D scenes from a single snapshot compressive imaging measurement, which folds many masked camera views into one picture. It uses a large vision model to guess geometry and poses from proxy views, then refines a Gaussian scene with a purpose-built densification rule.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on VGGT returning reliable poses and geometry from proxy views built from one multiplexed measurement; this out-of-distribution transfer is never validated against ground-truth-view or ground-truth-pose initialization.","rationale":"The reader's CONDITIONAL verdict is well supported. I independently identify the same load-bearing assumption: the entire geometry and pose stream originates from VGGT applied to synthetic proxies derived from one multiplexed measurement. The paper's controlled variants, run-to-run stability report, and explicit dynamic-scene limitation are genuine strengths, and I do not see an internal contradiction or any indication that the central claim is false. The secondary concern about choosing the Gaussian-count reference iteration (7,000) by test-set performance is real but would mainly inflate reported margins, not invalidate the architecture. The decisive missing experiment is the oracle comparison above. If proxy-based initialization matches ground-truth-view initialization, the paper should be accepted as stated; if not, the conditional verdict should require either an oracle study or a re-scoped claim. Because the requested check is absent, the current CONDITIONAL verdict remains appropriate.","tokens_in":22763,"tokens_out":4463,"duration_ms":49502,"concrete_test":"Run the identical coarse-stage protocol on the six standard scenes (Table I) and the six complex scenes (Table III) with three initializations: (A) proposed proxy views through VGGT; (B) ground-truth latent views X_i through the same frozen VGGT; (C) ground-truth poses and sparse geometry injected directly as an upper bound. Report PSNR, SSIM, LPIPS, and ATE after the same 20,000-iteration coarse stage for all three. If (A) is within a pre-registered margin of (B), e.g., at most 0.5 dB PSNR and 1% nATE, and (C) does not substantially exceed both, then the transfer concern is resolved. If (B) or (C) clearly outperforms (A), the headline result depends on an unvalidated VGGT generalization and the claim would need to be re-scoped or strengthened with an explicit oracle study.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B builds the entire pipeline on Phi_init(\\tilde X), where \\tilde X is a sequence of proxy views produced from a single SCI measurement by ENI or AMDI (Eq. 3). These proxies retain mask-selected samples and fill omitted pixels by interpolation or local-constant inversion, so they are not view-resolved RGB images and are far outside VGGT's training distribution. Yet the coarse-stage optimization (Eq. 5) and the final trajectory accuracy (Table VI) start from VGGT's pre-BA/post-BA/dense-fallback geometry and poses, and pose refinement is only a local SE(3) adjustment initialized at identity. If VGGT's output on these proxies is biased or scene-dependent, no downstream component can recover the true trajectory, and the reported 3D quality could be an artifact of an unmeasured generalization of a frozen model. The ablation in Table VII shows that removing VFM poses drops PSNR by 5.27 dB, confirming that the pose initialization is load-bearing, while Table XIII compares only proxy-construction strategies and never against ground-truth views or poses. The paper therefore does not yet establish that the single-measurement-to-3D claim is independent of VGGT's undocumented behavior on a synthetic input type.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GS2CI, a framework for reconstructing a static 3D Gaussian scene and the camera trajectory from a single snapshot compressive imaging (SCI) measurement and known binary masks. It constructs proxy views from the measurement via energy-normalized initialization (ENI) or adaptive mask-decoding initialization (AMDI), feeds them to the frozen 3D vision foundation model VGGT to obtain initial poses and sparse geometry, and then jointly optimizes 3D Gaussians and per-view SE(3) adjustments under an SCI forward-model loss augmented by a mean-opacity regularizer. A new densification rule, OSGR, augments split candidates using local opacity statistics and bounds Gaussian growth. After coarse convergence, a frozen 2D diffusion model (DiFix3D+) generates pseudo-view targets at interpolated and extrapolated poses for a weighted fine-stage loss. Experiments on simulated and real SCI data compare against GAP-TV, PnP methods, EfficientSCI, SCINeRF, SCIGS, controlled 3DGS variants, and a reproduced SCISplat, reporting PSNR/SSIM/LPIPS and trajectory metrics. The paper also reports run-to-run stability for the coarse stage and acknowledges that dynamic scenes remain a limitation.","tokens_in":22985,"tokens_out":6968,"duration_ms":70324,"significance":"If the central claim holds, the paper would be a significant step: single-measurement SCI-to-3D reconstruction with foundation-model priors, at about 68 minutes of training versus about 746 minutes for SCINeRF. The work has clear strengths: the controlled Gaussian variants (SCI-3DGS, SCI-MCMC, SCI-RevADC) isolate OSGR from the shared reconstruction framework, trajectory accuracy is evaluated against ground-truth poses, Table XII provides three-seed variance for the coarse stage, code is released, and the static-scene limitation is stated explicitly. However, the central claim is not fully established because the load-bearing VGGT initialization is applied to out-of-distribution proxy views without an oracle comparison, and one core OSGR hyperparameter (the Gaussian-count reference iteration) was selected on the main evaluation set. The contribution is promising but conditional on these points being addressed.","major_comments":[{"comment":"The paper's central claim depends on VGGT returning reliable poses and geometry from proxy views that are not view-resolved RGB images, yet this transfer is never validated against an oracle initialization. Eq. (3) constructs proxies from a single multiplexed measurement via ENI or AMDI, and Table XIII compares only proxy-construction strategies; it never compares against initialization from ground-truth latent views or ground-truth poses. Table VII shows this is load-bearing: removing VFM poses drops PSNR from 35.57 to 30.30 dB, and pose refinement in Section III-C is only an identity-initialized SE(3) adjustment, so a biased VGGT trajectory cannot be corrected downstream. I request an explicit experiment that initializes the same pipeline from ground-truth views and/or ground-truth poses, plus reporting of how often each initialization branch (PreBA, post-BA, dense fallback) is selected per scene; without this, the end-to-end success could rest on unmeasured generalization of a frozen model.","section":"Section III-B, Eq. (3)-(4); Table XIII; Table VII"},{"comment":"The Gaussian-count reference iteration is selected on the same six-scene set used for the main results. Appendix B states that \"we compare reference iterations of 5,000, 6,000, 7,000, 8,000, and 9,000. The 7,000-iteration setting gives the best overall result among these capped configurations and is fixed for all reported experiments.\" Because the count cap is a core component of OSGR and the six scenes are also the evaluation set of Tables I and VII, this is a test-set hyperparameter choice and can inflate the reported gains. The revision should select this iteration on a held-out validation split or report results for all candidate iterations on the test scenes, and should clarify how the claim that \"hyperparameters remain unchanged across standard and extended evaluations\" is compatible with this selection.","section":"Appendix B"},{"comment":"The comparison against SCISplat† is computed on only 17 of 20 cases because \"SCISplat† fails to initialize on three\" cases, but the paper does not report which scenes fail, the failure criteria, or whether the failure is due to the SCI input or to the reproduction itself. Since Table IX is the basis for the headline 3.29 dB PSNR improvement and the best-ATE claim, failure-dependent selection can bias the comparison. Please report the scene identities and failure reasons, provide results on all 20 cases under a predefined handling rule, and state whether the reproduction was checked against the published SCISplat outputs on any common scene.","section":"Table IX and Section IV-D"}],"minor_comments":[{"comment":"The captions state that bold and underlined values denote best and second-best results, but many entries in the tables appear with neither formatting; please ensure the final PDF renders the formatting consistently.","section":"Tables I and III"},{"comment":"The real-data evaluation is qualitative only; adding a quantitative metric on the real SCI data, even a limited one, would strengthen the claim that the method transfers beyond simulation.","section":"Section IV-B and Fig. 4"},{"comment":"The paper should add the static-scene qualifier to the abstract's single-measurement claim, since Table XI shows that SCI-MCMC outperforms the proposed method on both dynamic sequences and the limitation is acknowledged only later in the paper.","section":"Section IV-E and Abstract"},{"comment":"Three-seed variance is reported only for the coarse stage on six scenes; the main tables and the challenging-setting tables are single-seed, so please report variance or seed sensitivity for at least the key claims in Tables I, III, and VIII.","section":"Table XII and main tables"},{"comment":"The sentence \"for larger inputs, odd window sizes grow geometrically until full-image coverage\" is ambiguous because Eq. (18) gives a finite list; please specify exactly how the list is extended for larger images.","section":"Appendix A, Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is plausible, but the two principal concerns—unvalidated VGGT transfer on proxy views and test-set selection of the Gaussian-count reference iteration—are load-bearing and need additional experiments. Both are fixable in revision, so I do not recommend rejection. The revision should include the oracle-initialization study and a held-out validation protocol for OSGR hyperparameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nQuick take: this paper is a solid, well-engineered contribution to the SCI-to-3D literature, and the OSGR densification rule is the most defensible new idea in it. But the central pipeline rests on the assumption that VGGT, a model trained on ordinary view-resolved images, gives reliable poses and geometry when fed proxy views reconstructed from a single multiplexed measurement. That assumption is load-bearing and never directly tested. The paper deserves serious refereeing, but I would not let the current version through without an oracle-style comparison.\n\nWhat is actually new: OSGR—opacity-guided splitting plus mean-opacity regularization plus a count cap—is a genuine, mechanics-level fix for the instability of adaptive density control under SCI supervision. The controlled comparisons against SCI-3DGS, SCI-MCMC, and SCI-RevADC within the same framework do isolate OSGR properly, and the 3.29 dB gain over the SCISplat reproduction is reported honestly with the 17-case caveat. The authors are also candid about dynamic-scene failure, which I respect. The code being available is a plus.\n\nThe soft spots are real but fixable. First, the Gaussian-count reference iteration was chosen as the best among 5k–9k on the same six scenes that produce the headline numbers. That is mild test-set model selection; it probably does not change the qualitative story, but it should be acknowledged or cross-validated. Second, SCI-Gaussian [16] is cited but never compared. If it is the closest 3DGS-based SCI baseline, omitting it weakens the empirical claim. Third, and most important: the whole initialization is a frozen VGGT applied to ENI/AMDI proxies that are far outside its distribution. Table XIII shows adaptive routing helps, but it does not compare against an oracle that gives VGGT the ground-truth views or poses. Given that removing VFM poses costs 5.27 dB PSNR, the method's success depends on VGGT's out-of-distribution behavior, and that behavior is currently a black box.\n\nI do not think these issues sink the paper. The core idea—use geometry-oriented proxy views to unlock foundation-model priors for single-shot SCI reconstruction—is plausible and the empirical results are consistently strong. But the authors need to either validate VGGT's transfer on proxy views or show the method is robust to its failure. A comparison with GT-view initialization and GT-pose initialization, even as an upper bound, would answer the question directly. Until then, treat the headline numbers as conditional.\n\nWho is this for: anyone working in SCI, computational photography, or robust 3DGS initialization. It is worth a serious referee slot, and I would recommend major revision with the oracle experiment as the key request.","headline":"A well-engineered SCI-to-3D pipeline with a genuinely new densification rule, but the load-bearing VGGT-on-proxy initialization is never validated against ground-truth views or poses.","tokens_in":23624,"tokens_out":2080,"would_cite":true,"duration_ms":19637,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One snapshot compressive measurement, plus the known coding masks, suffices to recover a high-quality 3D scene and an accurate camera trajectory through foundation-model priors and an opacity-guided densification rule.","keywords":["snapshot compressive imaging","3D Gaussian splatting","vision foundation model priors","single-shot 3D reconstruction","novel view synthesis","camera pose estimation","opacity-guided densification","multi-view SCI"],"falsifier":"Take one benchmark scene, synthesize its SCI measurement, and run the full pipeline twice, once with the proxy views fed to the vision model and once with the same measurement but ground-truth poses and geometry as the initialization: a nearly identical result would show the claimed prior is not load-bearing, whereas a collapse would confirm the dependence. A sharper check is to feed the proxy pipeline a scene with repetitive texture or strong specularity and compare the vision model's pose estimates against the known trajectory before any Gaussian optimization; if the trajectory error spikes on such scenes, the end-to-end gains are hostage to the model's out-of-distribution behavior on invented proxy images.","tokens_in":22449,"feed_emoji":"📷","tokens_out":13162,"duration_ms":117494,"temperature":0.7,"pith_summary":"Snapshot compressive imaging folds many masked views into one sensor readout; under camera–scene motion during the exposure those views become different viewpoints, so a single measurement carries 3D information. The paper claims that this one measurement, together with the known coding masks, is enough to recover a high-quality static 3D scene and an accurate camera trajectory, provided the optimization starts from geometry and poses proposed by a large vision model and is kept stable by a custom opacity-guided densification rule. If true, a single coded exposure could replace a camera array for multi-view capture, and the recovered 3D scene would explain the measurement without ever seeing an unmixed view. The paper reports consistent gains over existing 3D-from-SCI methods across standard, unbounded, high-compression, and varied mask-ratio settings, while cutting per-scene training time from roughly twelve hours to about one.","feed_headline":"One coded snapshot reconstructs a scene and its camera path","feed_subtitle":"Foundation-model priors plus an opacity-aware density rule recover 3D geometry and poses from one fused exposure.","key_machinery":"The load-bearing mechanism is the pairing of a proxy-view initialization chain with a densification rule named OSGR (Opacity-Guided Splitting and Growth Regulation). Proxy construction turns the single measurement $Y$ into $N$ complete RGB grids: ENI normalizes each pixel by its mask multiplicity and interpolates mask-omitted locations when the average multiplicity is below a threshold, while AMDI solves small local least-squares systems for per-view contributions when the multiplicity is high. A frozen 3D vision foundation model maps those proxies to a sparse point set and initial poses, giving the later optimization a geometry to hold onto. OSGR then guards the Gaussian optimization against the degeneracy that a single multiplexed residual can be explained by raising opacity instead of moving geometry: each Gaussian with $\\alpha_i > \\bar{\\alpha}_i$, where $\\bar{\\alpha}_i$ is the mean opacity of its three nearest neighbors, enters the split-candidate set; the loss carries a mean-opacity penalty; extra split candidates are capped at five percent of the population; and the population at iteration 7,000 becomes a hard upper bound for all later splitting and duplication. A frozen 2D diffusion prior adds the final stage, proposing pseudo-view targets at interpolated and extrapolated poses whose influence is weighted by accumulated $\\alpha$.","core_discovery":"The paper's central claim is that the snapshot compressive imaging equation $Y = \\sum_{i=1}^N X_i \\odot M_i + Z$, the coded sum of $N$ masked views plus noise, can be inverted jointly for a static 3D Gaussian scene and for the $N$ camera poses, rather than merely for a stack of 2D frames. Because the masks are known but the sum destroys view identity, the authors supply structure from outside the measurement: a feed-forward 3D vision foundation model ingests proxy views (energy-normalized when mask overlap is low, locally decoded by regularized least squares when overlap is high) and returns initial sparse geometry and poses, which are then refined by optimizing the Gaussian scene directly against the coded sum while pose corrections are learned in the Lie algebra. The paper's own contribution to stability is OSGR, which adds to the split candidate set any Gaussian whose opacity exceeds the mean opacity of its nearest neighbors, penalizes global mean opacity in the loss, caps additional candidates, and freezes the Gaussian count after a reference iteration, preventing the optimizer from absorbing the multiplexed residual by inflating opacity instead of moving geometry. In the reported experiments this configuration gives the best or near-best reconstruction quality on six standard scenes, the best coarse-stage novel-view quality, the best results across mask ratios from 0.125 to 0.75 and compression ratios 8, 16, and 32, the lowest trajectory error among the compared methods, and a 3.29 dB PSNR gain with a 32-minute training saving over the authors' own reproduction of the closest published pipeline. The paper also reports that dynamic scenes remain beyond the static formulation, where a Markov-chain densification baseline scores higher on two video sequences.","pith_inferences":["The reported headline margins come from the first-stage reconstruction in the controlled comparisons, so they measure the proxy initialization plus OSGR, not the diffusion-based refinement; the auxiliary 2D-prior stage contributes only small additional gains in the component study.","An oracle experiment replacing the proxy-derived poses with ground-truth poses would quantify how much of the success is the vision model's generalization to out-of-distribution proxy images; the paper does not run it, and the answer would determine whether the method transfers to scenes unlike its benchmarks.","The same recipe, build proxy views that respect a known forward model, let a large pretrained model propose geometry, then optimize a differentiable scene against the true measurement, should carry over to rolling-shutter video, motion-blurred captures, or coded-aperture systems wherever one measurement multiplexes several viewpoints.","The dynamic-scene failure and the paper's own opacity-peak explanation suggest a concrete extension: tag Gaussians whose local opacity contrast persists after several densification steps, and route only those into a motion-aware representation such as 4D Gaussians, which would test whether the static-scene assumption is the actual bottleneck."],"forward_implications":["A single coded exposure taken during camera–scene motion becomes a substitute for a multi-camera array: the measured sum plus the known masks carries enough multi-view information to reproduce both the static 3D scene and the camera trajectory.","The recovered scene supports novel-view synthesis at unseen poses, with the auxiliary 2D prior contributing mainly at extrapolated viewpoints where the measurement itself offers the least guidance.","SCI reconstruction shifts from decoding a stack of 2D frames to jointly recovering a 3D scene and poses, at a training cost of roughly one hour per scene rather than the twelve-plus hours reported for the NeRF-based baseline.","The ablations attribute the stability to the OSGR rule and pose refinement specifically: removing the opacity regulation or the count cap degrades quality or exhausts memory, so the gains do not come from the foundation-model initialization alone.","The improvements persist across mask ratios from 0.125 to 0.75 and compression ratios of 8, 16, and 32, which the paper takes as evidence that the approach is not tuned to a single acquisition regime."],"supporting_citations":[{"why":"The 3D vision foundation model that converts the proxy-view sequence into initial sparse geometry and camera poses; the entire pipeline starts from its output.","marker":"[17]"},{"why":"Defines the Gaussian scene representation and differentiable rasterizer whose adaptive density control OSGR replaces for the SCI setting.","marker":"[12]"},{"why":"The NeRF-based SCI-to-3D baseline across all comparisons and the source of the real SCI data used for evaluation.","marker":"[14]"},{"why":"The earlier 3DGS-based SCI baseline that anchors the comparison table for reconstruction from a snapshot.","marker":"[13]"},{"why":"The frozen 2D vision model (DiFix3D+) that generates pseudo-view targets for the auxiliary fine-stage refinement.","marker":"[30]"},{"why":"Supplies the transmittance-preserving opacity correction applied to duplicated and split Gaussians inside OSGR.","marker":"[40]"},{"why":"The closest published pipeline, reproduced by the authors on 17 scenes to establish the 3.29 dB PSNR and training-time comparison.","marker":"[15]"},{"why":"The energy-normalization idea that the low-multiplicity proxy construction (ENI) adapts to compensate for mask-overlap exposure variation.","marker":"[3]"}],"fun_headline_variants":["One coded snapshot yields 3D scene and camera path","Single measurement to 3D: Gaussian splatting and VFM priors","Coded snapshot to 3D: scene geometry and poses in one shot","Snapshot compressive imaging decodes to 3D with foundation models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that a vision model trained on ordinary photos will return reliable camera poses and geometry even when its input is a sequence of proxy views stitched or inverted out of one multiplexed measurement, and the paper does not compare that initialization against ground-truth views or poses; if the model misreads those proxies, the scene and trajectory have nothing else to anchor them.","fun_headline_variants_meta":{"raw":{"variants":["One coded snapshot yields 3D scene and camera path","Single measurement to 3D: Gaussian splatting and VFM priors","Coded snapshot to 3D: scene geometry and poses in one shot","Snapshot compressive imaging decodes to 3D with foundation models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000284,"raw_usage":{"total_tokens":1787,"prompt_tokens":1169,"completion_tokens":618,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":785,"completion_tokens_details":{"reasoning_tokens":541}},"tokens_in":785,"tokens_out":618,"duration_ms":6414,"temperature":1.0,"reasoning_tokens":541,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:28:32.767823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one benchmark scene, synthesize its SCI measurement, and run the full pipeline twice, once with the proxy views fed to the vision model and once with the same measurement but ground-truth poses and geometry as the initialization: a nearly identical result would show the claimed prior is not load-bearing, whereas a collapse would confirm the dependence. A sharper check is to feed the proxy pipeline a scene with repetitive texture or strong specularity and compare the vision model's pose estimates against the known trajectory before any Gaussian optimization; if the trajectory error spikes on such scenes, the end-to-end gains are hostage to the model's out-of-distribution behavior on invented proxy images.","supporting_citations":[{"cited_title":"VGGT: Visual geometry grounded transformer,","cited_arxiv_id":null,"evidence_quote":"The 3D vision foundation model that converts the proxy-view sequence into initial sparse geometry and camera poses; the entire pipeline starts from its output."},{"cited_title":"3D gaussian splatting for real-time radiance field rendering,","cited_arxiv_id":null,"evidence_quote":"Defines the Gaussian scene representation and differentiable rasterizer whose adaptive density control OSGR replaces for the SCI setting."},{"cited_title":"SCINeRF: Neural radiance fields from a snapshot compressive image,","cited_arxiv_id":null,"evidence_quote":"The NeRF-based SCI-to-3D baseline across all comparisons and the source of the real SCI data used for evaluation."},{"cited_title":"SCIGS: 3D gaussians splatting from a snapshot compressive image,","cited_arxiv_id":null,"evidence_quote":"The earlier 3DGS-based SCI baseline that anchors the comparison table for reconstruction from a snapshot."},{"cited_title":"DIFIX3D+: Improving 3D reconstructions with single-step diffusion models,","cited_arxiv_id":null,"evidence_quote":"The frozen 2D vision model (DiFix3D+) that generates pseudo-view targets for the auxiliary fine-stage refinement."},{"cited_title":"Revising densification in gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Supplies the transmittance-preserving opacity correction applied to duplicated and split Gaussians inside OSGR."},{"cited_title":"Learning Radiance Fields from a Single Snapshot Compressive Image","cited_arxiv_id":"2412.19483","evidence_quote":"The closest published pipeline, reproduced by the authors on 17 scenes to establish the 3.29 dB PSNR and training-time comparison."},{"cited_title":"BIRNAT: Bidirectional recurrent neural networks with adversarial training for video snapshot compressive imaging,","cited_arxiv_id":null,"evidence_quote":"The energy-normalization idea that the low-multiplicity proxy construction (ENI) adapts to compensate for mask-overlap exposure variation."}],"review_version":1}