{"id":"8f224b64-bbab-4936-a8d5-188a9d7d041d","arxiv_id":"2507.18344","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A real-time RGB-D SLAM system that represents surfaces as 2D Gaussian disks and uses them in GICP tracking plus a normal-supervised loss, improving reconstruction accuracy on Replica and TUM-RGBD.","lead":"This paper builds a real-time SLAM system that maps rooms with flat, surface-hugging Gaussian disks instead of puffy 3D blobs, which cuts depth errors sharply. It also feeds surface normals into the optimization, improving geometric accuracy while keeping tracking fast.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline reconstruction gain rests on an unexplained, order-of-magnitude-discrepant reproduction of GS-ICP SLAM (20.54 cm Depth L1) that conflicts with the paper's own baseline (4.18 cm in Table V), so the central comparison is not yet trustworthy.","rationale":"The reader identified the singular GICP covariance as the weakest assumption. I agree that Section III-C is underspecified and Eq. (12) has a sign error, but this is not the load-bearing point: even a perfectly regularized covariance would leave the central comparative claim standing or falling on the baseline. The paper's own Table V appears to contradict Table I: the 'Baseline' row has PSNR 37.92, identical to the GS-ICP SLAM row in Table III, while its Depth L1 is 4.18 cm--five times smaller than the 20.54 cm reported for the 'reproduced' official GS-ICP SLAM. If the Baseline is GS-ICP SLAM, the external reproduction is implausible; if it is not, the comparison table is not labeled clearly enough to support a '96% improvement' claim. This is an internal-consistency problem, not a matter of disagreeing with field consensus. The concrete test is cheap and decisive: rerun the official baseline with the same evaluation pipeline. The covariance singularity, by contrast, admits standard fixes such as pseudo-inverse or epsilon regularization and would not invalidate the internal ablation. For that reason I partially agree with the reader's overall CONDITIONAL verdict but not with the choice of weakest assumption. This may be a metric or evaluation-protocol mismatch rather than a defective method, but the paper must disclose and correct it before the central claim is established.","tokens_in":11915,"tokens_out":6578,"duration_ms":68246,"concrete_test":"Obtain the released official GS-ICP SLAM code, run it on all 8 Replica sequences using the exact TSDF fusion and metric script used for Table I, and report per-scene Depth L1, Precision, Recall, and F1. Then state explicitly what Table V's 'Baseline' row is. If the official baseline is close to 4.18 cm (Table V) rather than 20.54 cm (Table I), the headline improvement over GS-ICP SLAM is an artifact of an unfair reproduction; if it is genuinely 20.54 cm, provide the reconstructed meshes and the full command and evaluation configuration so the discrepancy can be audited.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest quantitative support for the central reconstruction claim is Table I, where G2S-ICP SLAM gets 0.74 cm Depth L1 and GS-ICP SLAM* gets 20.54 cm. This baseline number is internally inconsistent: Table V, whose 'Baseline' has the same PSNR (37.92 dB) as the GS-ICP SLAM row in Table III, reports a Depth L1 of 4.18 cm for the starting point. If that baseline is GS-ICP SLAM, the two tables cannot both be right; if it is a different 3D-Gaussian system, the caption does not say so and the official-code reproduction remains unexplained. Either way the headline margin (claimed 96% improvement) is computed against a number that the paper itself does not reproduce. No code or evaluation script is provided, so one cannot tell whether the 20.54 cm figure comes from a different TSDF integration, a different depth mask, an untuned hyperparameter setting, or an actual failure mode of the official release. The secondary technical issues--the singular covariance S=diag(s1,s2,0) without epsilon in Section III-C, and the sign error in the GICP objective in Eqs. (11)-(12)--are real but would not change the need for a fair external comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes G2S-ICP SLAM, an RGB-D Gaussian Splatting SLAM system that represents the scene with surface-aligned 2D Gaussian disks instead of volumetric 3D ellipsoids. The 2D disks are integrated into a Generalized ICP (GICP) tracker through an anisotropic planar covariance prior, and the mapping is supervised by photometric, depth, and normal consistency losses, together with a distance-aware scale initialization. The system is evaluated on the Replica and TUM-RGBD datasets, reporting real-time operation at about 30 FPS, a 0.74 cm average Depth L1 on Replica, competitive ATE, and slightly lower rendering quality compared to its main baseline, GS-ICP SLAM.","tokens_in":12272,"tokens_out":3640,"duration_ms":36293,"significance":"If the claims hold, the paper addresses a real limitation of volumetric 3DGS-based SLAM, namely multi-view depth inconsistency, by enforcing a surface-aligned representation within both tracking and mapping. The use of external benchmarks, ablations, and quantitative comparisons against several recent baselines is a strength, as is the explicit reporting of runtime. However, the central quantitative claims currently rest on an internally inconsistent baseline reproduction and on an underspecified tracking formulation, so the contribution cannot be fully validated from the manuscript as written.","major_comments":[{"comment":"The reported comparison against the main baseline is internally inconsistent. Table I lists GS-ICP SLAM* with an average Depth L1 of 20.54 cm, while Table V, whose 'Baseline' row has the same PSNR (37.92 dB) as the GS-ICP SLAM* row in Table III, reports a rendered Depth L1 of 4.180 cm. If the baseline in Table V is GS-ICP SLAM, the two tables cannot both be correct; if it is a different 3D-Gaussian system, the caption and text do not identify it. This discrepancy directly affects the claimed 96% depth improvement over the main baseline and must be resolved with a consistent evaluation protocol.","section":"Table I vs. Table V"},{"comment":"The maximum-likelihood derivation of the GICP objective has a sign error. The Gaussian log-likelihood is proportional to -d^T Sigma^{-1} d, so Eq. (12) should minimize (or maximize the negative of) the squared Mahalanobis distance, not maximize the positive term as written. As stated, the objective would prefer arbitrarily large residuals. Please correct the equation and verify that the implemented tracking objective matches the corrected form.","section":"Eq. (11)-(12), Section III-B"},{"comment":"The 2D disk covariance S = diag(s1, s2, 0) is singular, yet the GICP formulation in Eqs. (9)-(12) requires the inverse of C_src and C_tgt. The paper does not specify any epsilon regularization, pseudo-inverse, or rank-deficient handling. Furthermore, the text says the 2D anisotropic covariance is embedded into each C_src_m, but the construction of C_src from the depth-map neighborhood (Section III-B) and the exact substitution with the 2D disk covariance are not given. This missing detail is load-bearing for the claimed tracking improvement and must be clarified.","section":"Section III-C, Eq. (1), Eqs. (9)-(12)"},{"comment":"The full model's depth accuracy is reported differently in two tables. Table I gives G2S-ICP SLAM an average Depth L1 of 0.74 cm on Replica, while Table V reports the full model's 'Rendered Depth L1' as 0.437 cm. If these are the same metric, the discrepancy needs explanation; if they are different metrics (e.g., rendered depth versus TSDF-integrated mesh depth), the definitions should be stated explicitly in the evaluation section.","section":"Table I vs. Table V, Section V-E"}],"minor_comments":[{"comment":"There are many typographical errors, e.g., 'Exisiting', 'aprroaches', 'incorportate', 'Generailzed', 'transfromation', 'awrae', 'coparisons', and 'incredible' in Table IV's caption. A careful proofread is needed.","section":"Throughout"},{"comment":"The 2D Gaussian kernel is written as G(q) = exp(u^2+v^2 / 2), which lacks the negative sign; it should be exp(-(u^2+v^2)/2).","section":"Eq. (4)"},{"comment":"The alpha term in the depth rendering equation is written as 'alpha G_j(x)' instead of 'alpha_j G_j(x)', consistent with Eq. (5).","section":"Eq. (6)"},{"comment":"There is a stray period in 'lambda_d L_d .' before the plus sign; the equation should read L = lambda_p L_p + lambda_d L_d + lambda_GAN L_GAN.","section":"Eq. (17)"},{"comment":"The phrase 'achieves the lowest tracking performance' should be 'achieves the lowest tracking error'; otherwise it contradicts the following sentence.","section":"Section V-C"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic, but the central comparison relies on a baseline reproduction whose numbers conflict within the manuscript itself. If the authors can provide a corrected, consistent evaluation and clarify the covariance regularization, the contribution may be salvageable; however, without these fixes the main claims are not verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Key take: the core idea—putting 2D Gaussian disks into a GICP tracker with an anisotropic covariance prior—is legitimately new, and the internal ablation shows it works. But the headline comparison against GS-ICP SLAM is built on a baseline number (20.54 cm Depth L1) that conflicts with the paper's own ablation baseline (4.18 cm) at identical PSNR, so the 96% improvement claim doesn't hold up yet. The GICP derivation also has a sign error and the zero-variance covariance is left singular without regularization.\n\nWhat's actually good: the integration is not trivial; using the same 2D Gaussian representation for both tracking and mapping is a sensible design choice, and the normal-consistency loss is a natural fit. The ablation on Replica is clean: depth L1 goes from 4.18 to 2.08 with the disk representation, then to 0.44 with the geometry-aware loss. That is credible internal evidence that the components do something. The system runs at 30 FPS, matching GS-ICP SLAM, and the qualitative figures show visibly cleaner meshes. The paper is clearly written and the related work is fair.\n\nThe soft spots: first, the external comparison. Table I says GS-ICP SLAM* (official code) gets 20.54 cm, but Table V's baseline—which has the same PSNR as GS-ICP SLAM in Table III—gets 4.18 cm. The asterisk doesn't explain that. One of these numbers is wrong, or the two tables aren't measuring the same thing. The paper never tells you which. That's a load-bearing flaw for the main claim. Second, Eqs. (11)-(12) write the GICP objective as an argmax of the squared Mahalanobis distance; MLE minimizes it. The surrounding text says 'minimizing', so it's probably a typo, but it needs fixing. Third, the paper says S=diag(s1,s2,0) for all Gaussians, making the covariance singular, and never mentions adding a small epsilon to keep the Mahalanobis distance invertible. That may just be an implementation detail omitted for brevity, but it's the kind of thing a reviewer needs to see.\n\nThis paper deserves a serious referee, but not in its current form. The baseline must be re-run or the discrepancy explained, and the GICP math needs to be made correct and concrete. If that happens, the result would be a solid within-subfield contribution for dense RGB-D SLAM. I'd send it to review, but I'd ask the authors to be very specific about the baseline.","headline":"Plausible surface-aligned Gaussian SLAM idea undercut by an unexplained baseline discrepancy and a shaky GICP derivation; worth refereeing after the numbers are fixed.","tokens_in":12721,"tokens_out":3114,"would_cite":false,"duration_ms":29593,"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":"Modeling every map point as a flat 2D disk aligned with the local surface lets a Gaussian-Splatting RGB-D SLAM system reconstruct geometry far more accurately and completely in real time, while keeping pose tracking competitive.","keywords":["2D Gaussian Splatting","RGB-D SLAM","Generalized ICP","camera tracking","surface reconstruction","depth consistency","geometry-aware loss","real-time SLAM"],"falsifier":"On a Replica or TUM sequence, compute the eigenvalues of the combined covariance $C_{\\mathrm{tgt}} + T C_{\\mathrm{src}} T^T$ for every correspondence during GICP tracking and record the condition number. If any matrix is singular or near-singular and replacing $S=\\operatorname{diag}(s_1,s_2,0)$ with $S=\\operatorname{diag}(s_1,s_2,10^{-6})$ leaves ATE and Depth L1 essentially unchanged, the core claim that zero normal variance drives the improvement is falsified. A second check: run the full method with the geometry-aware normal loss removed ($\\lambda_{\\mathrm{GAN}}=0$) and compare Depth L1; if it jumps back toward 2 cm, most of the reconstruction gain comes from the loss, not the disk representation.","tokens_in":11754,"feed_emoji":"🗺️","tokens_out":7643,"duration_ms":63049,"temperature":0.7,"pith_summary":"This paper claims that the geometric shape of the Gaussian primitive is the bottleneck in Gaussian-Splatting SLAM: volumetric 3D ellipsoids let each surface point be seen at different depths from different views, while a 2D Gaussian disk constrained to the local tangent plane removes that ambiguity. It embeds this anisotropy into a Generalized ICP tracker by giving every source point a covariance with zero variance along the surface normal, and it supervises mapping with photometric, depth, and surface-normal losses. On Replica the reported average Depth L1 falls to 0.74 cm, versus 20.54 cm for the reproduced GS-ICP SLAM baseline, with higher precision and recall while running at 30 FPS. The upshot is that real-time dense mapping can have surface-consistent geometry without sacrificing photorealistic rendering.","feed_headline":"Flattening map points into disks cuts SLAM depth error 28-fold","feed_subtitle":"A real-time Gaussian Splatting SLAM aligns each primitive with the local surface plane, boosting geometry without slowing tracking.","key_machinery":"The load-bearing object is the 2D Gaussian disk, a surface element with covariance $C_k = R_k S_k S_k^T R_k^T$, where $R_k=[t_1,t_2,n_k]$ and $S_k=\\operatorname{diag}(s_1,s_2,0)$. Its role is to force each primitive's spatial uncertainty onto the local tangent plane, eliminating depth ambiguity along the normal; the same covariance is used both in rendering and, as an anisotropic prior on source points, inside the Generalized ICP Mahalanobis-distance registration. A distance-aware initialization $s_1,s_2 \\propto 1/z^p$ and a geometry-aware normal loss complete the mechanism.","core_discovery":"The central claim is that surface-aligned 2D Gaussian disks, not volumetric 3D ellipsoids, should be the primitive of a real-time Gaussian Splatting SLAM system. Each Gaussian is parameterized by a center $p_k$, a rotation $R_k=[t_1,t_2,n_k]$ built from two tangent vectors and the surface normal, and a scale matrix $S_k=\\operatorname{diag}(s_1,s_2,0)$, so its support collapses onto the local tangent plane. The same covariance, with zero variance along the normal, is injected into each source point's distribution in the Generalized ICP registration, so tracking and mapping share one geometric prior. A geometry-aware loss then supervises rendered color, depth, and surface normals, and the scale is initialized from sensor depth so pixels far away grow larger Gaussians. The result, the paper argues, is a map that is geometrically accurate and complete while pose estimation remains at least as accurate as the 3D-ellipsoid baselines.","pith_inferences":["Editorial inference: if the zero-normal covariance is regularized in the GICP distance, the reported tracking gains may persist with better numerical conditioning; a direct ablation with $S=\\operatorname{diag}(s_1,s_2,\\epsilon)$ would separate the geometric prior from the singularity concern.","Editorial inference: the same tangent-plane covariance could be applied to monocular or stereo Gaussian-Splatting SLAM where depth is estimated, but the zero-variance normal becomes harder to justify when depth is uncertain; testing on real noisy sensors would show how much of the Replica gain transfers.","Editorial inference: the geometry-aware normal loss may be the main driver of the depth-L1 improvement, since the ablation shows the 2D-disk switch alone reduces depth error from 4.18 cm to 2.08 cm while the full optimization brings it to 0.44 cm; isolating these two contributions on TUM-RGBD would clarify the mechanism.","Editorial inference: applying the same normal and depth supervision to volumetric 3DGS SLAM without the disk constraint could test whether the flat primitive or the supervision is responsible, offering a clean control experiment."],"forward_implications":["If the central claim holds, Gaussian-Splatting SLAM can produce depth-accurate surface meshes directly, with reported Replica Depth L1 under 1 cm, rather than relying on post-hoc TSDF fusion to clean up volumetric artifacts.","Tracking does not pay a price for the geometric prior: average ATE on Replica is reported at 0.15 cm, slightly better than the 3D-ellipsoid baseline, so surface constraints can serve as a regularization signal rather than a handicap.","Reconstruction completeness improves substantially (F1 81.57% on Replica vs 6.04% for the reproduced GS-ICP baseline), so the approach addresses holes and missing structures in real-time maps.","Rendering quality drops slightly (PSNR 36.88 dB vs 37.92 dB for the baseline) but stays competitive with other real-time GS SLAM systems, meaning the geometric gains come at a modest photometric cost.","Because the GICP formulation is unchanged and only the covariance prior is injected, the method slots into existing RGB-D GS-ICP pipelines without architectural redesign."],"supporting_citations":[{"why":"Supplies the 2D Gaussian disk representation whose tangent-plane constraint is the central primitive.","marker":"[14]"},{"why":"Provides the Generalized ICP registration algorithm into which the anisotropic covariance prior is injected.","marker":"[15]"},{"why":"The main baseline; the paper extends it by replacing 3D ellipsoids with 2D disks and compares against it.","marker":"[10]"},{"why":"Defines 3D Gaussian Splatting rendering, the representation the paper argues is geometrically inconsistent.","marker":"[8]"},{"why":"SplaTAM, a comparative real-time Gaussian Splatting SLAM baseline for tracking and rendering.","marker":"[11]"},{"why":"MonoGS, a comparative Gaussian Splatting SLAM baseline used in tracking and rendering tables.","marker":"[12]"},{"why":"GS-SLAM, a comparative dense Gaussian Splatting SLAM baseline used in reconstruction evaluation.","marker":"[13]"},{"why":"Replica dataset, the primary benchmark for reconstruction, tracking, and rendering experiments.","marker":"[24]"},{"why":"TUM RGB-D dataset, the real-world benchmark for tracking and rendering evaluation.","marker":"[25]"}],"fun_headline_variants":["2D Gaussian disks sharpen SLAM geometry and tracking","Surface-aligned Gaussians boost SLAM accuracy in real time","G2S-ICP SLAM flattens points into disks for precise tracking","Anisotropic Gaussians improve SLAM rendering and pose","Real-time Gaussian Splatting SLAM with planar primitives"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that setting every Gaussian's variance along the surface normal to exactly zero in the GICP covariance stays numerically stable and that this zero normal variance is what improves tracking; the paper specifies no regularization term to keep the Mahalanobis distance invertible.","fun_headline_variants_meta":{"raw":{"variants":["2D Gaussian disks sharpen SLAM geometry and tracking","Surface-aligned Gaussians boost SLAM accuracy in real time","G2S-ICP SLAM flattens points into disks for precise tracking","Anisotropic Gaussians improve SLAM rendering and pose","Real-time Gaussian Splatting SLAM with planar primitives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000727,"raw_usage":{"total_tokens":3260,"prompt_tokens":951,"completion_tokens":2309,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":2222}},"tokens_in":567,"tokens_out":2309,"duration_ms":15524,"temperature":1.0,"reasoning_tokens":2222,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:13:39.142058+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a Replica or TUM sequence, compute the eigenvalues of the combined covariance $C_{\\mathrm{tgt}} + T C_{\\mathrm{src}} T^T$ for every correspondence during GICP tracking and record the condition number. If any matrix is singular or near-singular and replacing $S=\\operatorname{diag}(s_1,s_2,0)$ with $S=\\operatorname{diag}(s_1,s_2,10^{-6})$ leaves ATE and Depth L1 essentially unchanged, the core claim that zero normal variance drives the improvement is falsified. A second check: run the full method with the geometry-aware normal loss removed ($\\lambda_{\\mathrm{GAN}}=0$) and compare Depth L1; if it jumps back toward 2 cm, most of the reconstruction gain comes from the loss, not the disk representation.","supporting_citations":[{"cited_title":"2d gaussian splat- ting for geometrically accurate radiance fields,","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D Gaussian disk representation whose tangent-plane constraint is the central primitive."},{"cited_title":"Rgbd gs-icp slam,","cited_arxiv_id":null,"evidence_quote":"The main baseline; the paper extends it by replacing 3D ellipsoids with 2D disks and compares against it."}],"review_version":2}