{"id":"246853bf-b574-4692-a9f2-06e3e143e54c","arxiv_id":"2412.06250","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Splatter-360 is an end-to-end generalizable 3D Gaussian splatting model that builds a spherical cost volume to improve geometry and rendering from wide-baseline panoramic images.","lead":"Splatter-360 is a neural network that creates new views of a scene from two wide-baseline 360-degree photos, using 3D Gaussian splatting with a spherical cost volume to estimate depth. It is designed for VR and simulation users who want fast rendering from sparse panoramic captures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline comparison is confounded: Splatter-360 is trained with ground-truth depth supervision and DepthAnythingV2 monocular features, while MVSplat is trained with RGB loss only, so the reported PSNR and depth gains may reflect extra supervision rather than the proposed spherical cost volume.","rationale":"The reader's weakest_assumption identifies the supervision mismatch, and I agree this is the most load-bearing issue. The central claim is not merely that Splatter-360 is a strong system; it asserts that spherical cost volume construction improves geometry estimation. Because the method includes depth supervision and monocular depth priors, the key controlled variable is missing. A secondary concern is internal consistency: the 'Full' model in Table 3 (PSNR 29.121 on Replica, 27.487 on HM3D) does not match Table 1 (29.888 and 28.293), suggesting the ablations may use a different training budget; this weakens the ablation attributions but does not by itself refute Table 1. I would keep the conditional verdict: the synthesis-quality claim is plausible and the margins over PanoGRF are large, but the architecture attribution and the fairness of the depth comparison require the proposed controlled retraining before the paper can be fully accepted.","tokens_in":15131,"tokens_out":8159,"duration_ms":77128,"concrete_test":"Retrain both models on HM3D under identical supervision: (a) RGB+LPIPS loss only for both; (b) add depth loss with weight 0.1 for both; (c) add frozen DepthAnythingV2 features for both. Use the same optimizer, iterations, batch size, and number of GPUs, and evaluate on the Replica split with Table 1 PSNR/SSIM/LPIPS and Table 2 Abs Rel/delta<1.25. If Splatter-360's margin over MVSplat persists under (c), or equivalently if removing depth loss and mono features from Splatter-360 still leaves it ahead, the supervision confound is resolved; if the margin shrinks below about 0.3 dB or reverses, the headline superiority cannot be attributed to the proposed architecture.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Table 1: PSNR 28.293 vs 27.179 on HM3D; 29.888 vs 28.399 on Replica) and the geometry claim (Table 2) compare Splatter-360 against MVSplat under asymmetric training signals. Splatter-360 is trained with an explicit depth loss of weight 0.1 (Sec. 4.1) and feeds frozen DepthAnythingV2 monocular depth features into its bi-projection encoder (Sec. 3.1, Eq. 1); MVSplat is trained with RGB loss only. The paper never ablates the depth loss, and the only monocular-feature ablation (Table 3, 'Mono Feat.') retains the depth loss. Consequently, the reported 1.1-1.5 dB PSNR lead and the depth-metric lead may be caused by the additional supervision and monocular prior rather than by the spherical cost volume or bi-projection encoder. The abstract attributes the gain to spherical-domain matching ('enhancing depth perception and geometry estimation'), so this confound directly threatens the paper's causal claim, not just the raw ranking.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Splatter-360, an end-to-end generalizable 3D Gaussian splatting framework for novel view synthesis from wide-baseline 360-degree panoramic images. The method constructs a spherical cost volume via a spherical sweep algorithm, uses a bi-projection encoder that combines equirectangular and cubemap features with pretrained monocular depth features, incorporates cross-view attention, and predicts pixel-aligned Gaussians. The authors report evaluations on HM3D and Replica against PanoGRF, MVSplat, DepthSplat, and HiSplat, claiming consistent improvements in synthesis quality and depth estimation. The supplementary material contains additional ablations, including an ablation of a depth refinement U-Net that is not described in the main text.","tokens_in":15367,"tokens_out":6699,"duration_ms":60526,"significance":"If the central comparisons hold, this is a strong empirical contribution. The spherical cost volume is a natural adaptation of plane-sweep matching to panoramic imagery and avoids behind-camera sampling artifacts, and the prospect of real-time rendering from sparse panoramic captures is practically relevant. The paper releases code and trained models, and the ablations (e.g., Supplementary Table 6) show that the refinement components matter. However, the current evidence for the central causal claim, namely that spherical-domain matching drives the gains, is weakened by a supervision confound in the comparison with MVSplat and by an internal inconsistency between the main results and the ablation baseline.","major_comments":[{"comment":"The full model in the ablations does not reproduce the headline model. Table 1 reports Splatter-360 at 28.293 dB PSNR on HM3D and 29.888 dB on Replica, while the 'Full' rows of Table 3 and Supplementary Table 6 report 27.487 dB and 29.121 dB on the same datasets, differences of 0.806 dB and 0.767 dB. Because all ablation deltas are computed against this different baseline, the module-contribution claims in Section 4.5 are not quantitatively supported as stated. Please explain the discrepancy (e.g., different number of GPUs, batch size, or training budget) or retrain the ablations to match the model used in Table 1.","section":"Table 1 vs Table 3 and Supplementary Table 6"},{"comment":"The depth and geometry comparison with MVSplat is confounded by asymmetric training signals. Splatter-360 is trained with an explicit depth loss of weight 0.1 (Sec. 4.1) and consumes monocular depth features from a pretrained DepthAnythingV2 (Sec. 3.1, Eq. 1), while MVSplat is trained with an RGB loss only. The depth-metric lead in Table 2 and the qualitative depth advantage in Fig. 4 could therefore be caused by the additional supervision or monocular prior rather than by the proposed spherical cost volume or bi-projection encoder. The '× Mono Feat.' ablation in Table 3 removes the monocular features but retains the depth loss, so it does not isolate the cause. Please ablate the depth loss, or train MVSplat with the same depth supervision, and report the resulting PSNR and depth metrics.","section":"Sec. 4.1, Sec. 3.1 (Eq. 1), Table 2"},{"comment":"The main-text method description omits the depth refinement U-Net. Section 3.2 describes only the spherical cost volume and its refinement U-Net (Eq. 7), yet Fig. 1 and Supplementary Sec. 7.3 state that a separate depth refinement U-Net exists, and Supplementary Table 6 shows that removing it costs about 0.7 dB PSNR on both datasets. The architecture evaluated in Tables 1-4 is therefore not fully described in the main text. Please include this component in Section 3.2 and in the ablation study, or explicitly refer the reader to the supplementary description at the point where the refinement U-Net is introduced.","section":"Sec. 3.2 and Supplementary Sec. 7.3/Table 6"}],"minor_comments":[{"comment":"The sentence 'To proveide a fair comparisions' contains typos; it should read 'To provide a fair comparison'.","section":"Sec. 4.2"},{"comment":"'Corss-view attention' should be 'Cross-view attention'.","section":"Sec. 4.5"},{"comment":"The caption says the depth comparison is with PanoGRF, but Section 4.4 and the surrounding text describe a comparison with MVSplat; please align the caption with the method actually shown.","section":"Fig. 4"},{"comment":"The near/far settings are described confusingly: the text first says near=0.5 and far=10 for the perspective baselines, then says near=0.1 was also tried and gave worse results; please clarify which configuration produced the numbers in Table 1.","section":"Supplementary Sec. 7.2"},{"comment":"The paper claims real-time rendering but reports no runtime or throughput measurement anywhere; please add a quantitative runtime comparison (e.g., rendering fps on a specified GPU) to substantiate this claim.","section":"Abstract and Conclusion"},{"comment":"For the models trained by the authors (MVSplat† and Splatter-360†), please report the number of training iterations, batch size, learning rate, and the exact train/validation split so that the comparisons are reproducible.","section":"Sec. 4.3/Supplementary"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency between Table 1 and the 'Full' row of the ablations is more than a presentational issue; if the ablation baseline is not the same model as the one in the headline comparison, the ablation study needs to be re-run or clearly justified. The supervision confound with MVSplat should also be addressed directly, either by ablating the depth loss in Splatter-360 or by giving MVSplat the same depth supervision. Both issues are fixable within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: the spherical cost volume for feed-forward panoramic 3DGS is a real contribution, and the internal ablations back it up, but the headline comparison against MVSplat is not clean. Splatter-360 trains with ground-truth depth supervision (weight 0.1) and monocular depth features from DepthAnythingV2; MVSplat† is trained with RGB only. The paper never ablates the depth loss, so part of the reported 1.1–1.5 dB PSNR lead, and most of the depth-metric lead, could come from the extra supervision rather than from spherical matching.\n\nWhat is new and good: matching directly in the spherical domain, via a spherical sweep cost volume, is a sensible adaptation of MVSplat’s plane sweep to 360° imagery. The paper gives a concrete reason why the naive cube-map conversion is problematic: 3D sample points in a planar cost volume can project behind the camera in a source view, and ERP avoids that. The bi-projection encoder with cross-view attention is a reasonable extension of existing 360° feature extractors, and the fusion of monocular depth features is well motivated for textureless indoor regions. The ablations in Table 3 are compelling: removing the spherical cost volume costs 5.27 dB on Replica and 2.26 dB on HM3D, and removing cross-view attention or the ERP branch also hurts. That is solid internal evidence that the proposed modules matter.\n\nSoft spots, in proportion: the missing depth refinement U-Net is the biggest process issue. Supplementary Table 6 says it contributes about 0.7 dB PSNR on both datasets; that is not a tiny trick, and the main text should not omit it. Second, the comparison with MVSplat† in Tables 1 and 2 is asymmetric. Even if the RGB-only training is standard for generalizable 3DGS, the paper’s claim that the spherical cost volume “enhances depth perception and geometry estimation” is not established against MVSplat unless the depth loss is ablated. Third, no error bars or multiple seeds are reported; given the small Replica test set (90 trajectories), this is a minor but real limitation.\n\nThe citation pattern looks fine; the self-citation of DiffPano for the spherical projection formula is legitimate. The code and models are promised, which is good.\n\nVerdict: this paper deserves a serious referee. The core idea is new for the panoramic setting, the method is well engineered, and the internal evidence is strong. The authors should be asked for a depth-loss ablation, a complete architecture description, and ideally a retrained MVSplat† with depth supervision to isolate the contribution of spherical matching. I would bring it to a reading group and would cite it if I worked in panoramic NVS.","headline":"The spherical cost volume is a real contribution with strong internal ablations, but the headline lead over MVSplat is confounded by asymmetric depth supervision and needs an ablation.","tokens_in":15936,"tokens_out":3402,"would_cite":true,"duration_ms":32454,"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":"Performing multi-view matching directly in the spherical domain is the key to generalizable 3D Gaussian splatting for wide-baseline panoramic images.","keywords":["generalizable 3D Gaussian splatting","panoramic novel view synthesis","wide-baseline 360° images","spherical cost volume","spherical sweep algorithm","bi-projection encoder","cross-view attention","equirectangular and cubemap projection"],"falsifier":"Retrain MVSplat (or remove depth supervision from Splatter-360) with the supervision held fixed: give MVSplat the same depth loss and the same monocular depth features while keeping its cubemap plane-sweep cost volume, and compare PSNR and depth metrics on HM3D. If the margin over MVSplat collapses, the paper's evidence does not establish that spherical-domain matching is the cause.","tokens_in":14891,"feed_emoji":"🌐","tokens_out":11571,"duration_ms":96988,"temperature":0.7,"pith_summary":"Splatter-360 is a feed-forward 3D Gaussian splatting model that takes two (or three) wide-baseline 360° panoramas and renders novel panoramic views in real time. Its central claim is that multi-view matching should be done directly on the spherical image domain: a spherical sweep algorithm samples depth candidates along rays and builds a spherical cost volume, instead of flattening each panorama into cubemaps and running plane-sweep matching, which can sample points behind the source camera. Combined with a bi-projection encoder that fuses equirectangular and cubemap features under cross-view attention, and with monocular depth features injected as geometry priors, the network predicts pixel-aligned Gaussian primitives on the equirectangular grid. On HM3D and Replica, the paper reports PSNR of 28.29 and 29.89 dB, beating the strongest perspective baseline (MVSplat trained on panoramas) by 1.11 and 1.49 dB and the panoramic NeRF baseline PanoGRF by 2.66 and 1.97 dB, with better depth estimates. The reason this matters is that wide-baseline panoramic captures are cheap to acquire but hard to render into; if the claim holds, real-time high-quality view synthesis from sparse 360° images is within reach for VR and simulation pipelines.","feed_headline":"Spherical sweep gives 360° splatting a 1.1-1.5 dB edge","feed_subtitle":"Feed-forward model turns sparse 360° panoramas into real-time novel views.","key_machinery":"The spherical cost volume is the central object. For each reference pixel, depth candidates $r_m$ are sampled logarithmically between $r_{\\text{near}}$ and $r_{\\text{far}}$, converted from equirectangular coordinates $(\\theta,\\phi)$ to Cartesian camera coordinates, back-projected into each source view with the relative pose, and compared by feature dot product; softmax over the $D=128$ candidates yields a spherical depth estimate that directly determines Gaussian centers. This is what lets the network avoid the behind-camera ($z<0$) sampling problem of cubemap plane sweeps. Around it sit the bi-projection encoder, which fuses equirectangular and cube-map features with cross-view attention and monocular depth features, and a refinement U-Net that corrects the cost volume.","core_discovery":"On its own terms, the paper claims that a generalizable 3D Gaussian splatting pipeline can work end-to-end on wide-baseline panoramas if it matches features in spherical coordinates. The central comparison is that Splatter-360 outperforms the state-of-the-art perspective generalizable splatting methods MVSplat, DepthSplat, and HiSplat, as well as the panoramic radiance-field method PanoGRF, on HM3D and Replica across PSNR, SSIM, LPIPS, and estimated-depth metrics. The authors attribute the gain to three design choices: the spherical cost volume from the spherical sweep, the 3D-aware bi-projection encoder with cross-view attention, and the injection of monocular depth features. Ablations show that removing the spherical cost volume costs 5.27 dB on Replica and 2.26 dB on HM3D, while removing cross-view attention costs about 0.90 dB on Replica and 0.57 dB on HM3D, which supports the claim that spherical-domain matching carries most of the improvement.","pith_inferences":["The paper does not isolate the supervision confound; a fair test would retrain MVSplat with the same depth loss and monocular features to see whether the spherical cost volume is the real cause.","Because matching runs at 1/8 resolution with 128 depth bins, the same spherical sweep should extend to more input views or higher resolutions; the paper only reports two-view and three-view results.","The behind-camera problem it identifies is not limited to panoramas, so the spherical sweep idea could transfer to fisheye or wide field-of-view perspective rigs.","The pose requirement is the main practical gap; a pose-refinement head trained jointly with the spherical cost volume would be the natural step toward pose-free 360° reconstruction."],"forward_implications":["If the central claim is right, feed-forward panoramic novel view synthesis from two wide-baseline inputs becomes a real-time rasterization problem rather than a slow NeRF-style optimization.","The reported depth gains (Abs Rel 0.063 vs 0.088 on Replica and 0.076 vs 0.094 on HM3D, with $\\delta<1.25$ above 94%) imply that spherical-domain matching improves geometry estimation from sparse 360° views.","The method generalizes across datasets: after training on HM3D it keeps a clear margin on held-out Replica scenes with no per-scene optimization.","Ablation numbers imply the spherical cost volume, not the decoder, is the main source of the quality jump, so future panoramic splatting designs should focus on matching strategy.","The stated limitations mean the approach still assumes known poses, cannot generate unseen content, and is demonstrated only on indoor scenes."],"supporting_citations":[{"why":"MVSplat is the perspective plane-sweep generalizable Gaussian splatting baseline that Splatter-360 beats; it also supplies the refinement U-Net design and the D=128 depth sampling.","marker":"[5]"},{"why":"PanoGRF is the generalizable spherical radiance field baseline specialized for wide-baseline panoramas, the main NeRF competitor.","marker":"[8]"},{"why":"HM3D provides the training and test scenes for the main experiments (800 training, 100 test scenes).","marker":"[27]"},{"why":"Replica provides the held-out generalization test set of 18 indoor scenes.","marker":"[32]"},{"why":"DepthAnythingV2 supplies the pretrained monocular depth network whose features are injected into the cube-map branch.","marker":"[47]"},{"why":"DiffPano provides the spherical projection formula used to convert equirectangular coordinates to spherical and Cartesian coordinates in the spherical sweep.","marker":"[48]"},{"why":"UniMatch provides the convolutional backbone used for local feature extraction in both equirectangular and cubemap branches.","marker":"[44]"}],"fun_headline_variants":["Spherical matching powers real-time 360° splatting","Spherical cost volume beats perspective for 360° views","Wide-baseline panoramas get real-time novel views","Sparse 360° panoramas feed-forward to real-time views","Generalizable splatting handles sparse panoramas"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison with MVSplat is fair enough to attribute the gains to the spherical cost volume, even though Splatter-360 is trained with ground-truth depth supervision (loss weight 0.1) and pretrained monocular depth features that MVSplat does not receive.","fun_headline_variants_meta":{"raw":{"variants":["Spherical matching powers real-time 360° splatting","Spherical cost volume beats perspective for 360° views","Wide-baseline panoramas get real-time novel views","Sparse 360° panoramas feed-forward to real-time views","Generalizable splatting handles sparse panoramas"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001441,"raw_usage":{"total_tokens":5887,"prompt_tokens":1103,"completion_tokens":4784,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":719,"completion_tokens_details":{"reasoning_tokens":4701}},"tokens_in":719,"tokens_out":4784,"duration_ms":30076,"temperature":1.0,"reasoning_tokens":4701,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:51:42.480645+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain MVSplat (or remove depth supervision from Splatter-360) with the supervision held fixed: give MVSplat the same depth loss and the same monocular depth features while keeping its cubemap plane-sweep cost volume, and compare PSNR and depth metrics on HM3D. If the margin over MVSplat collapses, the paper's evidence does not establish that spherical-domain matching is the cause.","supporting_citations":[{"cited_title":"Panogrf: generalizable spherical radiance fields for wide-baseline panoramas","cited_arxiv_id":null,"evidence_quote":"PanoGRF is the generalizable spherical radiance field baseline specialized for wide-baseline panoramas, the main NeRF competitor."},{"cited_title":"Unifying flow, stereo and depth estimation","cited_arxiv_id":null,"evidence_quote":"UniMatch provides the convolutional backbone used for local feature extraction in both equirectangular and cubemap branches."}],"review_version":1}