{"id":"ed1381d8-bfcc-403d-bb7d-0c2cfe53f36b","arxiv_id":"2411.12471","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SCIGS is the first 3D Gaussian Splatting method that reconstructs an explicit dynamic 3D scene from a single compressed image, using camera pose stamps and a transformation network.","lead":"SCIGS reconstructs a 3D Gaussian Splatting scene from a single compressed snapshot image, handling both static and dynamic scenes. It swaps direct camera pose optimization for a network that transforms Gaussian primitives per camera pose stamp, aiming for multi-view consistent high-speed 3D capture.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3D reconstruction claim is untested: all reported metrics are same-view SCI decoding, and the Sec. 3.3 camera/Gaussian equivalence is never validated by novel-view or geometry experiments.","rationale":"The reader's weakest assumption—that the camera/Gaussian equivalence in Sec. 3.3 is untested—is essentially the same concern I reach, though I would frame it slightly more broadly: the training objective and all evaluation are single-view, so the method's central claim of explicit 3D reconstruction is not demonstrated by any multi-view or geometry evidence. This is a missing-support issue rather than an internal inconsistency: the pipeline could plausibly learn a coherent scene, but the paper does not show it. The proposed concrete test of held-out novel-view rendering directly settles whether the recovered Gaussians represent a genuine 3D scene or merely fit the coded measurement from the fixed viewpoint. Because this is fixable with additional experiments and not a demonstrated fatal flaw, the appropriate verdict remains CONDITIONAL, matching the reader's assessment.","tokens_in":12821,"tokens_out":5444,"duration_ms":58929,"concrete_test":"Add a held-out novel-view experiment on the static NeRF-Synthetic Hotdog scene and the LLFF Airplants scene. After training with the fixed camera, freeze the Gaussians and transformation network, render at 5-10 known held-out camera poses (e.g., translating the camera roughly 5% of the scene size or rotating it 5-10 degrees), and compare PSNR/SSIM/LPIPS against ground-truth frames. Also compare against the same test using EfficientSCI or SCINeRF outputs as 2D images. If the novel-view PSNR is within about 2 dB of the training-view PSNR and clearly better than a per-frame 2D baseline, the 3D claim survives; if novel-view rendering degrades sharply or is no better than baseline, the fixed-camera equivalence has not produced a real 3D representation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that SCIGS reconstructs an explicit 3D scene, but the optimization (Sec. 3.3, Eq. 7) never uses a second viewpoint: all B rendered frames are projected from one fixed camera, and the loss is only on the mask-modulated sum (Eq. 5). The stated equivalence between moving the camera and moving the Gaussians holds only if the per-stamp transformation is a single global rigid transform; the MLP outputs independent per-Gaussian (δμ, δr), so nothing enforces rigidity, and no experiment checks whether the learned transforms actually form a coherent 3D scene. In dynamic scenes the physical camera does not move, so the 'camera pose stamp' is just a frame index and the recovered per-frame deformations are not constrained by any multi-view consistency. Tables 1, 2, A, and B report only same-view image metrics, not novel-view synthesis, pose accuracy, or geometry. Thus the primary novelty—explicit 3D reconstruction from one compressed image—is not actually validated; the evidence is consistent with the model fitting the coded 2D measurement by per-frame deformation or hallucination.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SCIGS, a 3D Gaussian Splatting framework for reconstructing a 3D scene from a single snapshot compressive image. The method initializes random 3D Gaussians and a fixed camera, then uses an MLP transformation network that takes Gaussian coordinates and a camera pose stamp to output per-Gaussian position and quaternion increments. The transformed Gaussians are rendered through a high-frequency filter and a differentiable rasterizer into B frames, which are modulated by the known masks to synthesize a compressed image; training minimizes the loss between the synthesized and real compressed images. Experiments compare decoded frame quality against ground-truth frames on static (LLFF, NeRF Synthetic, DeblurNeRF) and dynamic (DAVIS2017) datasets, reporting PSNR/SSIM/LPIPS relative to existing SCI decoders and SCINeRF.","tokens_in":13090,"tokens_out":3462,"duration_ms":37472,"significance":"If validated, the paper would introduce the first explicit 3D representation reconstructed from a single snapshot compressive image, with a plausible route to dynamic scenes and a useful high-frequency filter for stabilizing Gaussian transformations. The dynamic-scene results are substantially better than SCINeRF on several datasets (e.g., Roundabout PSNR 31.07 vs. 26.02 in Table 2), and the filter ablation in Table 4 shows a clear quantitative benefit. However, the paper's central claim of 3D reconstruction is not directly tested: all reported metrics are same-view decoded-image metrics, and no experiment measures novel-view synthesis, geometry, or recovered camera poses. The contribution is therefore interesting and potentially important, but the evidence as presented does not yet establish that the method reconstructs a coherent 3D scene rather than fitting the coded 2D measurement with per-frame deformations.","major_comments":[{"comment":"The central claim of explicit 3D reconstruction is not validated by the experiments. All quantitative results in Tables 1, 2, A, and B compare the decoded frames against ground-truth frames at the same viewpoint used during training; there is no novel-view synthesis, no depth or geometry evaluation, and no recovered camera pose accuracy. Since the training loss (Eq. 5) is only the mask-modulated sum from one fixed camera, the reported metrics are equally consistent with a 2D per-frame decoder that uses a Gaussian parameterization. A minimal validation would be to render from a held-out camera pose (available in the static benchmarks) and compare against the corresponding ground-truth view, or to compare recovered depth or pose against known values.","section":"Section 3.3 and Section 4.2"},{"comment":"The equivalence asserted in Section 3.3 between moving the camera to the correct pose and moving the Gaussians in front of a fixed camera is only valid if the per-stamp transformation is a single global rigid transform. However, Eq. (7) outputs independent per-Gaussian increments (δμ, δr), and nothing in the loss or network architecture enforces rigidity or even consistency across Gaussians; the only smoothness argument is the continuity of the MLP. No experiment checks whether the learned per-stamp transformations actually form a coherent camera motion or a coherent scene. In dynamic scenes, the stamps are merely frame indices and the physical camera does not move, so the 'camera pose stamp' terminology does not correspond to any recovered pose. This leaves the mechanism that distinguishes SCIGS from a 2D SCI decoder untested.","section":"Section 3.3, Eq. (7)"},{"comment":"The claim that SCIGS 'outperforms or approaches' existing methods is not supported on all static scenes. On Airplants, SCIGS reports SSIM 0.7267 and PSNR 27.18, whereas EfficientSCI reports SSIM 0.9425 and PSNR 30.13, and SCINeRF reports SSIM 0.9335; the Airplants LPIPS of 0.3003 is far worse than every baseline. The paper's explanation (lack of texture and large camera movements) is plausible but is not quantified, and it weakens the general claim of superiority in static scenes. The authors should either restrict their claims to the scenes where the method is competitive or provide additional analysis of why these two scenes fail.","section":"Table 1 and Section 4.2"},{"comment":"The mask overlapping rate OR is selected empirically after observing test performance in Table 3, and the high-frequency filter scale gamma is never reported. Because OR is part of the forward model used to synthesize the training compressed image, choosing OR based on the evaluation datasets is a form of test-set leakage that can inflate reported performance. The value of gamma in Eq. (11) must be reported for reproducibility, and the authors should clarify whether OR and gamma were tuned on a held-out validation set or on the same test scenes used in Tables 1 and 2.","section":"Section 4.3 and Eq. (11)"}],"minor_comments":[{"comment":"There are several typos and grammar issues, including 'conpressive' in the Conclusion, 'embead' in Eq. (7), 'solution' instead of 'resolution' in Section 4.1, 'Tranformer' in the Related Work, and 'SOTA SCI image method (SCINeRF)' in the caption of Figure 5, where SCINeRF is not an SCI image method.","section":"Abstract and Section 1"},{"comment":"The implementation details are incomplete: the paper does not report the number of training iterations, learning rates, Gaussian count, or the value of the filter hyperparameter gamma. This information is needed to reproduce the results.","section":"Section 4.1"},{"comment":"The sentence 'the quality of reconstruction increases first and then decreases with the overlapping rate ranging from 0.125 to 0.75' is unclear because Table 3 shows a monotonic decrease from 0.25 to 0.75 after an increase from 0.125 to 0.25; the wording should be revised.","section":"Section 4.3"},{"comment":"The units and roles of gamma and v_i in the filter scale gamma / v_i are not fully explained; since v_i is a sampling frequency, the combination implies a dimensional scale that should be clarified to avoid ambiguity.","section":"Eq. (11)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising idea and the dynamic-scene results are strong, but the central 3D-reconstruction claim is currently supported only by same-view decoding metrics. The requested additions of novel-view or geometry experiments and a corrected evaluation protocol for OR/gamma are substantial but feasible within the scope of a revision. I would also encourage the authors to clarify the relationship of their transformation network to SCINeRF's pose optimization and to report pose accuracy explicitly if they claim to recover camera poses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zixu and Hao have written a paper that does something genuinely new: it builds an explicit 3D Gaussian representation from a single snapshot compressive image, and it works better than existing SCI decoders on dynamic scenes. The transformation network with camera pose stamps is a sensible way to avoid camera-pose optimization, and the high-frequency filter ablation shows a real artifact problem and a real fix. Tables 2 and A give convincing same-view PSNR/SSIM/LPIPS gains over SCINeRF and EfficientSCI on most dynamic scenes. That much is solid.\n\nBut the paper oversells the 3D part. The stress-test note is right: all reported metrics are SCI decoding at the same fixed viewpoint. There is no novel-view synthesis, no geometry evaluation, no recovered camera pose check. The loss (Eq. 5) only compares the mask-modulated sum of B rendered frames to the compressed measurement, all from one camera. Nothing in the optimization forces the per-stamp transformed Gaussians to form a coherent 3D scene. The Sec. 3.3 equivalence between moving the camera and moving the Gaussians holds only for a global rigid transform, but Eq. 7 outputs independent per-Gaussian deltas; nothing enforces rigidity. So the explicit 3D scene is a nice story, but the evidence is consistent with the model being a strong 2D decoder that uses per-frame deformation to fit the coded measurement. The phrase 'novel-view images' in the contributions is not backed by a single experiment.\n\nThe static-scene results are also mixed: SCIGS loses to SCINeRF and sometimes EfficientSCI on Airplants and Hotdog, and the authors explain this in Section 4.2 in an honest way. The mask overlapping rate and filter gamma are chosen empirically, without error bars—minor, but it would help to see variance.\n\nOverall: take the paper as an SCI decoding method and it is a good one, especially for dynamic scenes. Take it as a 3D reconstruction method and it is unverified. I would send it to peer review, but the reviewers should demand novel-view rendering, a pose or geometry sanity check, and ideally a multi-view consistency test. If those fail, the 3D framing needs to be dropped and the paper repositioned as a high-quality SCI decoder.","headline":"SCIGS is a plausible and novel SCI decoder with strong dynamic-scene numbers, but its central claim of explicit 3D reconstruction is untested since every metric is same-view decoding.","tokens_in":13602,"tokens_out":2885,"would_cite":true,"duration_ms":27882,"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":"A single snapshot compressed image can reconstruct an explicit 3D Gaussian scene, including dynamic scenes, via a camera-pose-aware transformation network.","keywords":["snapshot compressive imaging","3D Gaussian splatting","dynamic scene reconstruction","camera pose estimation","transformation network","high-frequency filter","SCI decoding","novel view synthesis"],"falsifier":"Render the reconstructed Gaussians from a novel camera pose, not the fixed one, and compare against ground-truth images from a dataset with known camera poses, such as the LLFF or NeRF Synthetic scenes used in the paper. If the novel-view PSNR is close to that of a random 3D point cloud, the equivalence between camera motion and Gaussian transformation would be refuted; the paper reports no such test.","tokens_in":12625,"feed_emoji":"📸","tokens_out":4582,"duration_ms":38743,"temperature":0.7,"pith_summary":"This paper attempts to establish that one snapshot compressed image, of the kind a coded-aperture camera captures, contains enough information to reconstruct an explicit 3D scene as a set of 3D Gaussians, including when the scene is dynamic. It proposes SCIGS, a 3D Gaussian Splatting variant that replaces camera-pose optimization with a learned transformation of Gaussian primitives conditioned on camera pose stamps. The paper reports that this method surpasses prior SCI decoders and SCINeRF on dynamic scenes, and matches or exceeds them on most static scenes. If true, it would mean high-speed dynamic 3D scenes could be recovered from a single exposure with ordinary camera hardware.","feed_headline":"One compressed snapshot reconstructs a dynamic 3D scene","feed_subtitle":"It is the first 3D Gaussian Splatting method to decode one coded-aperture exposure, and it leads on dynamic scenes.","key_machinery":"The load-bearing object is the camera-pose-aware transformation network $F$, an MLP that maps a Gaussian's positional embedding and a camera pose stamp to $(\\delta\\mu, \\delta r)$. It does the work that camera-pose optimization would do, but per primitive, so different stamps can move the same base Gaussians differently, which is what lets a single compressed exposure decode into a dynamic scene. The correctness of this substitution rests on the assertion that moving the initial camera to the true pose is equivalent to moving the Gaussians in front of a fixed camera. A high-frequency filter, implemented by convolving each 3D Gaussian with a low-pass Gaussian whose scale depends on the maximal sampling frequency across all stamps, removes aliasing artifacts.","core_discovery":"SCIGS is the first method to recover an explicit 3D Gaussian Splatting scene from a single snapshot compressive image. The central mechanism is a primitive-level transformation network that takes each Gaussian's coordinates and a camera pose stamp as input and outputs increments to position and quaternion, effectively moving the Gaussians to the positions they would occupy for that camera pose. Because the initial camera is never moved, the method sidesteps the chaotic gradients that arise when most Gaussians project outside the correct image region. A high-frequency filter, derived from the Nyquist-Shannon sampling theorem, suppresses artifacts from the transformed Gaussians. The paper demonstrates the approach on static and dynamic scenes, reporting that it surpasses prior works on dynamic datasets.","pith_inferences":["The paper does not validate the recovered camera pose or report novel-view synthesis, so its strongest implied claim, that the reconstructed Gaussians form a geometrically correct 3D scene, remains untested; a natural extension would be measuring multi-view consistency on held-out views.","The equivalence between camera movement and Gaussian transformation, if correct, suggests a general recipe for other inverse problems where camera poses are unknown: optimize the object instead of the observer.","Since the method conditions the transformation on camera pose stamps, it could in principle render arbitrary intermediate times between the compressed frames, effectively interpolating high-speed motion beyond the exposure's temporal resolution, something the paper does not examine.","For autonomous driving, a single snapshot compressive camera could replace a burst of exposures, but the method's 480p dynamic experiments and unvalidated geometry would need scaling tests first."],"forward_implications":["If the central claim is correct, a single coded-aperture exposure can yield an explicit 3D Gaussian scene rather than a stack of 2D images.","The transformation network can decouple the transformation field from the compressed image, allowing dynamic scenes to be reconstructed from one snapshot.","Replacing camera-pose optimization with Gaussian transformation removes the oscillation problem that occurs when most Gaussians are ineffective, making 3D reconstruction from a random initialization more stable.","The high-frequency filter provides a principled way to remove artifacts introduced when Gaussians are moved between pose stamps."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting representation and rasterization pipeline that SCIGS builds upon.","marker":"[13]"},{"why":"SCINeRF is the prior NeRF-based method for reconstructing 3D scenes from a single compressed image; SCIGS compares against it and must beat it on dynamic scenes.","marker":"[14]"},{"why":"Mip-Splatting's low-pass filtering inspires the high-frequency filter used to remove artifacts from transformed Gaussians.","marker":"[39]"},{"why":"EfficientSCI is the state-of-the-art deep learning SCI decoder that SCIGS compares against on both static and dynamic datasets.","marker":"[31]"},{"why":"COLMAP is invoked as the unavailable source of camera poses and point clouds, motivating the transformation-network substitution.","marker":"[27]"}],"fun_headline_variants":["First 3DGS from one compressed snapshot","Single coded image to dynamic 3D scene via SCIGS","SCIGS: first 3DGS from a snapshot compressive image","One compressed shot, dynamic 3D scene: SCIGS leads","From a single coded exposure to explicit 3D Gaussians"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method stands on the assumption that rearranging the 3D Gaussians in front of a fixed camera is exactly equivalent to moving the camera to its true pose, and that a smooth MLP, guided only by gradients from the few Gaussians that already project near the right place, can find that rearrangement without the camera pose ever being checked.","fun_headline_variants_meta":{"raw":{"variants":["First 3DGS from one compressed snapshot","Single coded image to dynamic 3D scene via SCIGS","SCIGS: first 3DGS from a snapshot compressive image","One compressed shot, dynamic 3D scene: SCIGS leads","From a single coded exposure to explicit 3D Gaussians"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00059,"raw_usage":{"total_tokens":2757,"prompt_tokens":920,"completion_tokens":1837,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":1749}},"tokens_in":536,"tokens_out":1837,"duration_ms":12960,"temperature":1.0,"reasoning_tokens":1749,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:28:26.763926+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the reconstructed Gaussians from a novel camera pose, not the fixed one, and compare against ground-truth images from a dataset with known camera poses, such as the LLFF or NeRF Synthetic scenes used in the paper. If the novel-view PSNR is close to that of a random 3D point cloud, the equivalence between camera motion and Gaussian transformation would be refuted; the paper reports no such test.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian Splatting representation and rasterization pipeline that SCIGS builds upon."},{"cited_title":"Scinerf: Neural radiance fields from a snapshot compressive image","cited_arxiv_id":null,"evidence_quote":"SCINeRF is the prior NeRF-based method for reconstructing 3D scenes from a single compressed image; SCIGS compares against it and must beat it on dynamic scenes."},{"cited_title":"Mip-splatting: Alias-free 3d gaussian splatting","cited_arxiv_id":null,"evidence_quote":"Mip-Splatting's low-pass filtering inspires the high-frequency filter used to remove artifacts from transformed Gaussians."},{"cited_title":"Efficientsci: Densely connected network with space-time factoriza- tion for large-scale video snapshot compressive imag- ing","cited_arxiv_id":null,"evidence_quote":"EfficientSCI is the state-of-the-art deep learning SCI decoder that SCIGS compares against on both static and dynamic datasets."},{"cited_title":"Structure-from-motion revisited","cited_arxiv_id":null,"evidence_quote":"COLMAP is invoked as the unavailable source of camera poses and point clouds, motivating the transformation-network substitution."}],"review_version":1}