{"id":"214ac5e2-04f0-44dc-82e6-b0f62e922d13","arxiv_id":"2507.20854","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A real-time RGB-D SLAM system built on 2D Gaussian surfels with a depth-distortion adaptive rendering rule and analytic surfel pose Jacobians reports accurate geometry and robust tracking on Replica, TUM-RGBD, and ScanNet++.","lead":"The paper builds a real-time RGB-D SLAM system that represents scenes with flat 2D Gaussian surfels instead of 3D ellipsoids, and renders depth and normals from those surfaces. This design is reported to reduce mapping artifacts and to track camera poses more reliably under large viewpoint changes than prior Gaussian-based SLAM.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The adaptive depth-distortion rule in Eq. 7 can systematically relabel background-edge pixels as foreground, so the claimed edge sharpening may be an artifact of the asymmetric D(x)>D_c(x) substitution; a single whole-scene Replica ablation cannot rule this out.","rationale":"I read the paper in good faith. The central claim has two pillars: 2D surfel adaptive mapping and surfel-based pose Jacobian tracking. The convergence-basin analysis is the strongest evidence for the tracking pillar, and the ScanNet++ results, if reproducible, are impressive. The most fragile link, however, is the adaptive rendering heuristic. It is the core novel mapping contribution, is supported by exactly one Replica ablation, and has a concrete failure mode: the conditional substitution D(x)>D_c(x) is asymmetric and biases toward the nearest max-weight surfel. At occlusion boundaries this can label true-background pixels as foreground, inflating edge sharpness in a way that whole-scene F1 and L1 metrics would not reveal. This concern is not about disagreement with consensus; it is a correctness risk internal to the proposed mechanism. I also considered the missing supplementary Jacobian derivation and the lack of a GauS-SLAM comparison, but those are verifiability/scope issues rather than a specific mechanism that could invalidate the claimed behavior. The proposed synthetic edge-band test would settle whether the adaptive rule actually preserves background surfaces. If it passes, the conditional concern is retired and the paper remains conditionally acceptable pending code, error bars, and the supplementary derivation. If it fails, the mapping claim is weakened, but because the issue is localized and addressable, a conditional verdict with a request for the additional experiment still seems appropriate rather than outright rejection.","tokens_in":12252,"tokens_out":8319,"duration_ms":103852,"concrete_test":"On a synthetic two-plane occlusion scene (foreground plane at 1 m, background wall at 3 m, known step edge), run the S3LAM rasterizer with the adaptive rule (Eqs. 6–7) and compute signed depth error separately for pixels whose ground-truth surface is the background within a 3-pixel band of the edge. If the mean signed error of those background-edge pixels is negative (biased toward the foreground) by more than roughly 2% of the depth range, and removing the D(x)>D_c(x) condition eliminates that bias, the heuristic is demonstrably corrupting background geometry rather than merely sharpening edges.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise of the adaptive mapping contribution (Sec. 4.1, Eqs. 6–7) is that high depth distortion D_d reliably flags unfinished geometry, and that replacing alpha-blended depth and normal with the max-weight-surfel values, only when D(x)>D_c(x), sharpens edges without corrupting planar regions. The conditional is asymmetric: substitution happens only when the dominant surfel is closer than the blended depth. At an occlusion boundary, a background pixel whose ray passes within the footprint of a foreground surfel can have that foreground surfel as the max-weight one; then D(x)>D_c(x) is true and the pixel is re-labeled as foreground. This expands foreground silhouettes by roughly the surfel footprint and biases the 'sharp edge' improvement reported in Table 6. The ablation reports whole-scene F1/L1 on one synthetic dataset with no error bars, so a localized foreground-inflation error of a few centimeters would be hidden while still affecting precision/recall. Since mapping and tracking are coupled, this also threatens the tracking-robustness claim if the tracking loss uses the same rendered depth. No formal analysis or second dataset is given for when the substitution is safe.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents S3LAM, an RGB-D SLAM system built on 2D Gaussian surfel splatting. The scene is represented by oriented 2D Gaussians, and mapping optimizes color, depth, and normal losses with an adaptive rendering rule that replaces alpha-blended depth and normal with the max-blend-weight surfel's values when a depth-distortion uncertainty term exceeds a threshold and the blended depth is farther. Tracking optimizes the SE(3) pose using an analytic Jacobian derived from the surfel splatting model, including a radial component that the authors argue improves convergence under large viewpoint changes. Experiments on Replica, TUM-RGBD, and ScanNet++ report state-of-the-art or competitive tracking and reconstruction metrics, together with a convergence-basin analysis and ablations of the adaptive mapping and pose Jacobian components.","tokens_in":12576,"tokens_out":5850,"duration_ms":65525,"significance":"The paper targets a genuine limitation of 3DGS-based SLAM: the lack of explicit surface orientation in both mapping and pose optimization. If the stated results hold, the oriented-surfel representation with the analytic pose Jacobian is a meaningful step forward, and the convergence-basin comparison is a concrete, falsifiable test that is more informative than average ATE alone. The adaptive rendering idea is simple and cheap. However, the empirical support is currently weakened by single-run metrics without error bars, a central derivation that is deferred to a supplementary appendix that is not provided, an ablation whose control condition does not match the main result table, and a one-sided substitution rule that may systematically bias occlusion boundaries. These issues prevent me from recommending acceptance in the current form.","major_comments":[{"comment":"The adaptive substitution rule is asymmetric and can systematically relabel occlusion-boundary pixels as foreground. Consider a background pixel whose ray passes through the footprint of a foreground surfel; if that foreground surfel has the largest blending weight, Eq. (7) selects its depth, and because the alpha-blended depth is a weighted average of foreground and background depths, the condition D(x) > D_c(x) holds, so the pixel is replaced by the foreground depth. This expands foreground silhouettes by roughly the surfel footprint and can bias the claimed edge-sharpening improvement in Table 6; since the tracking loss in Eq. (8) uses the same rendered depth, it can also affect pose optimization. The only ablation is a whole-scene evaluation on Replica with no error bars and no second dataset, so a localized error of a few centimeters would not be detected. Please add a per-pixel analysis at occlusion boundaries, evaluate on a second dataset, report error bars, and provide a formal or empirical characterization of when the one-sided rule is safe.","section":"§4.1, Eq. (6)–(7)"},{"comment":"The central tracking contribution relies on the analytic Jacobian of the tangent vectors, but the derivation is only referenced as 'supplementary material,' which is not included with this submission. Without this derivation, the claimed advantage of the radial gradient and the 'w/o radial' ablation in Table 7 cannot be checked. Please include the full derivation in the paper, or at minimum state the exact closed-form expressions used to compute the terms in Eq. (10).","section":"§4.2, Eq. (9)–(10)"},{"comment":"The ablation's 'Ours' row (0.68, 1.21, 1.43, 1.69 cm on S0, S1*, S1, S2) does not match the main ScanNet++ results in Table 5 (0.35, 0.42, 0.51, 1.11 cm). No explanation is given for the discrepancy; if the ablation uses a different protocol, fewer iterations, or a different keyframe setting, the comparison against 'w/o depth loss' and 'w/o radial' is not controlled. Please reconcile the two tables or specify precisely what differs.","section":"Table 7"},{"comment":"All tracking and reconstruction tables report single-run metrics without error bars. Several reported differences are small (e.g., Table 3 Replica average 0.38 vs. 0.39 for SplaTAM; Table 4 fr2/xyz 1.16 vs. 1.24), and ScanNet++ trajectories with teleportations and large rotations may be sensitive to initialization and random seeds. Please report the number of runs and standard deviations, or otherwise justify why single-run evaluation is sufficient for the claims.","section":"Tables 3–5"}],"minor_comments":[{"comment":"The first row of Table 6 appears misaligned: the 3DGS row reads Acc.=2.80, P.=3.71, Comp.=74.8, R.=70.1, F1=72.3, L1=2.21. Precision cannot be 3.71% and completion cannot be 74.8 cm; likely the intended values are Acc.=2.80, P.=74.8, Comp.=3.71, R.=70.1, F1=72.3, L1=2.21. Please fix the column alignment.","section":"Table 6"},{"comment":"The notation D_c(x) = arg max_ω_k z_k is an abuse of notation: arg max returns the surfel index, not a depth value. Define k* = argmax_k ω_k and then D_c = z_{k*}, N_c = t_w^{k*}.","section":"Eq. (7)"},{"comment":"There is a typo in the Baselines paragraph: 'performation' should be 'performance'.","section":"Experiments, Baselines"},{"comment":"The convergence-basin protocol would be easier to interpret if the authors reported the number of sampled initial poses per distance, the exact pose-optimization settings used for MonoGS, and error bars or confidence intervals on the success-rate estimates.","section":"§5.3, Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic and the core idea is plausible, but the current manuscript does not yet support its strongest claims. The missing supplementary derivation of the pose Jacobian is a desk-level concern: the central technical contribution is not verifiable from the submission. The one-sided adaptive substitution rule also deserves closer scrutiny because it can systematically bias occlusion boundaries. I would ask the authors to provide the full derivation, reconcile Table 7 with Table 5, and add error bars or repeated-run statistics before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before reading it: it's a competent 2D-Gaussian-surfel SLAM system claiming better geometry and tracking than 3DGS-based SLAM, and the strongest evidence is a convergence-basin experiment that actually tests the core pose-optimization claim. It is not a field rewriter, but it is a real contribution that deserves peer review.\n\nWhat's new: bringing 2DGS into an online SLAM pipeline is not itself new (concurrent GauS-SLAM does it), but the depth-distortion adaptive rendering rule and the analytic SE(3) Jacobian with the radial gradient are. The radial gradient is the most interesting idea: the claim that it widens the convergence basin is supported by Figure 7, where the success rate for far initial poses is visibly better than MonoGS. The ablation in Table 7 also shows removing the radial term wrecks tracking on ScanNet++, which gives the claim some teeth. The mapping ablation (Table 6) shows the adaptive rule improves F1 over mean/median depth, though the gain is modest.\n\nSoft spots, in rough order of importance. First, all tables report single runs with no error bars. The ScanNet++ results are dramatic (0.35–1.11 cm vs. SplaTAM's 70.93 cm on full S1), and that kind of claim needs variance information. Second, the tangent-vector Jacobian derivation is deferred to a supplementary that I couldn't see; that's a load-bearing derivation and it needs to be in the paper or verifiable. Third, GauS-SLAM is cited but never compared quantitatively, which makes the 'state-of-the-art' wording too strong. Fourth, the adaptive rule in Eq. 7 genuinely worries me: substitution only happens when D > Dc, i.e. when the max-weight surfel is closer than the blended depth. At an occlusion boundary, a background pixel whose dominant surfel is a foreground fragment gets re-labeled as foreground, expanding silhouettes by roughly the surfel footprint. That could inflate the edge-sharpening results in Table 6. The single Replica ablation with no error bars can't rule this out, and the paper gives no analysis of when the substitution is safe. This is a plausible concern, not a proven flaw—but it needs addressing.\n\nWho this is for: researchers in dense RGB-D SLAM and Gaussian-splatting tracking. A serious referee should take it, but the authors need to address the missing derivation, add error bars, compare to GauS-SLAM, and analyze the adaptive rule's failure modes. I'd send it back with a clear 'major revision' verdict rather than reject it.","headline":"Solid, relevant Gaussian-surfel SLAM with a genuinely useful convergence-basin analysis and a plausible radial Jacobian, but it overclaims 'state-of-the-art' and needs error bars, a real GauS-SLAM comparison, and a careful look at the asymmetric adaptive-rendering rule before I'd trust the numbers.","tokens_in":13073,"tokens_out":2390,"would_cite":false,"duration_ms":27681,"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":"S3LAM claims that replacing 3D Gaussian ellipsoids with oriented 2D Gaussian surfels makes RGB-D SLAM both geometrically more accurate and far more robust to large viewpoint changes, with a depth-distortion adaptive rendering rule and an…","keywords":["2D Gaussian surfels","RGB-D SLAM","camera pose tracking","adaptive surface rendering","depth distortion","convergence basin","3D reconstruction"],"falsifier":"Run the paper's convergence-basin protocol with the radial-gradient term zeroed out: if the success-rate gap versus 3DGS pose optimization does not collapse, the radial term is not what widens the basin. A second check: on a planar scene with simulated depth noise, measure whether the adaptive $D(x) > D_c(x)$ substitution ever corrupts flat regions, visible as a drop in reconstruction F1.","tokens_in":12064,"feed_emoji":"📐","tokens_out":6191,"duration_ms":65197,"temperature":0.7,"pith_summary":"The paper claims that replacing 3D Gaussian ellipsoids with oriented 2D Gaussian surfels in an RGB-D SLAM system yields both more geometrically accurate reconstructions and more robust camera tracking, especially when successive frames barely overlap. It proposes two mechanisms for this: an adaptive surface rendering rule that detects unfinished geometry with a depth-distortion term and substitutes the dominant surfel's depth and normal where uncertainty is high, and an analytic pose Jacobian on SE(3) that includes a radial gradient term pulling the estimated orientation into alignment with the reconstructed surface. If the claim holds, SLAM systems can keep tracking through abrupt rotations and teleportations that currently break 3DGS-based trackers, and produce meshes with sharper edges at interactive frame rates. The evidence includes an F1 of 91.9 on Replica geometry, ScanNet++ ATE as low as 0.35 cm, and a convergence-basin analysis showing a wider basin than a representative 3DGS baseline.","feed_headline":"2D surfels give SLAM sharper maps and wider tracking convergence","feed_subtitle":"Depth-distortion rendering plus an analytic pose Jacobian keep S3LAM on track through sudden rotations and tiny frame overlap.","key_machinery":"The central object is the 2D Gaussian surfel: a flat, oriented Gaussian primitive defined by a center $p_k$, two tangent vectors $t_u$ and $t_v$ with scales $s_u, s_v$, a normal $t_w = t_u \\times t_v$, an opacity $\\alpha_k$ and a color $c_k$. Rays intersect its plane to get local $(u,v)$ coordinates, and $\\alpha$ blending over these plane intersections produces color, depth, and normal images. Two mechanisms carry the argument: (1) an adaptive mapping rule that computes a depth-distortion term $D_d = \\sum_{i,j} \\omega_i \\omega_j |z_i - z_j|$ over each pixel's surfel stack, and when $D_d$ exceeds a threshold $\\tau$ replaces averaged depth and normal with the values of the max-weight surfel (only when the averaged depth is larger), sharpening edges without punching holes in planes; and (2) an analytic pose Jacobian on the Lie algebra of SE(3) that, unlike the 3DGS Jacobian, includes the radial gradient $\\partial \\hat{t}_r / \\partial T_{CW}$ of the center-to-intersection vector, giving the optimizer a rotational signal that aligns camera orientation with the reconstructed surface. The convergence-basin experiment measures how often pose optimization from perturbed initial poses converges within 1 cm to the target, and attributes the wide basin to this radial term.","core_discovery":"S3LAM's central discovery is that the geometric nature of the scene primitive matters for both mapping and tracking: oriented surfels carry an explicit normal and a well-defined ray-plane intersection, so depth rendering is no longer a weighted average of Gaussian centers but a blending of per-ray intersection points. The paper derives the full SE(3) Jacobian of this rendering (Eq. 10) and shows it contains an extra radial component, the gradient of the vector from the surfel center to the ray-surfel intersection point, which gives the pose optimizer a direct signal to rotate the camera so its rays hit surfaces at the correct location. On the mapping side, the paper introduces a depth-distortion term (Eq. 6) computed from the blending weights and depths; pixels whose distortion exceeds a threshold are rendered by the dominant surfel rather than by opacity-weighted averaging, and this substitution is applied only when the rendered depth exceeds the dominant-surfel depth, preserving planar regions. The combined system tracks poses on ScanNet++ sequences with very small inter-frame overlap where SplaTAM fails, and reconstructs Replica geometry with precision and recall balanced at an F1 of 91.9, at 8.12 FPS and 4.2 GB memory.","pith_inferences":["A stress test with simulated depth noise on predominantly planar scenes would establish how often the conditional substitution $D(x) > D_c(x)$ corrupts planar geometry, a failure mode the paper does not analyze.","The radial-gradient analysis suggests a general principle: any renderer whose depth is defined by ray-primitive intersections, rather than weighted primitive centers, inherits a rotational gradient that can act as a free orientation prior; applying the same derivation to other surface-aware primitives may reproduce the convergence gain.","Because the paper only evaluates RGB-D SLAM, the same Jacobian could be ported as an initialization prior for point-cloud registration or visual odometry loops, where the convergence-basin question is the same."],"forward_implications":["SLAM systems that adopt oriented surfel primitives can track through abrupt viewpoint changes—rotations and teleportations of the kind in ScanNet++—that force 3DGS-based trackers to fail or to be truncated to short sequences.","Geometric reconstruction quality (F1, accuracy, completion) on standard indoor benchmarks reaches parity with NeRF-based SLAM while running at 8 FPS and using a fraction of the memory, removing a main practical obstacle to real-time use.","The analytic radial-gradient Jacobian gives a principled route to widen pose convergence basins, which could extend to other splatting-based trackers and to initialization of registration problems in general.","The depth-distortion adaptive rendering suggests that treating per-pixel uncertainty during rasterization is a lightweight alternative to denser surfel stores for real-time reconstruction of sharp edges."],"supporting_citations":[{"why":"Supplies the 2D Gaussian surfel representation and the CUDA rasterizer that S3LAM revises.","marker":"(Huang et al. 2024)"},{"why":"SplaTAM is the 3DGS SLAM baseline that fails on the full ScanNet++ S1 sequence, the key comparison for tracking robustness.","marker":"(Keetha et al. 2024)"},{"why":"MonoGS is the 3DGS pose-optimization baseline used in the convergence-basin comparison and a main tracking baseline on TUM.","marker":"(Matsuki et al. 2024)"},{"why":"Origin of the depth-distortion term that the paper repurposes as the uncertainty measure for adaptive mapping.","marker":"(Barron et al. 2022)"},{"why":"RTG-SLAM provides the surfel addition/deletion criteria and the optional ICP tracking module.","marker":"(Peng et al. 2024)"},{"why":"Supplies the Lie-algebra framework used for deriving the SE(3) pose Jacobians.","marker":"(Solà, Deray, and Atchuthan 2021)"},{"why":"GS-SLAM provides adaptive expansion and coarse-to-fine tracking ideas used in the surfel management design.","marker":"(Yan et al. 2024)"},{"why":"KinectFusion supplies the ICP method and the depth-gradient normal computation used in the pipeline.","marker":"(Newcombe et al. 2011)"}],"fun_headline_variants":["Surfel splatting refines SLAM mapping and pose tracking","2D surfels drive sharper SLAM geometry and faster pose convergence","Surfel-centric SLAM: depth exactness plus reliable tracking","S3LAM: geometrically accurate SLAM via surfel splatting","Surfel splats make SLAM tracking converge on tiny overlaps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a thresholded depth-distortion value reliably flags unfinished geometry in real-time SLAM, so that swapping in the dominant surfel's depth and normal on those pixels sharpens edges without damaging flat regions.","fun_headline_variants_meta":{"raw":{"variants":["Surfel splatting refines SLAM mapping and pose tracking","2D surfels drive sharper SLAM geometry and faster pose convergence","Surfel-centric SLAM: depth exactness plus reliable tracking","S3LAM: geometrically accurate SLAM via surfel splatting","Surfel splats make SLAM tracking converge on tiny overlaps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000534,"raw_usage":{"total_tokens":2588,"prompt_tokens":985,"completion_tokens":1603,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":1511}},"tokens_in":601,"tokens_out":1603,"duration_ms":13903,"temperature":1.0,"reasoning_tokens":1511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:11:25.243693+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the paper's convergence-basin protocol with the radial-gradient term zeroed out: if the success-rate gap versus 3DGS pose optimization does not collapse, the radial term is not what widens the basin. A second check: on a planar scene with simulated depth noise, measure whether the adaptive $D(x) > D_c(x)$ substitution ever corrupts flat regions, visible as a drop in reconstruction F1.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 2D Gaussian surfel representation and the CUDA rasterizer that S3LAM revises."},{"cited_title":"M.; Yang, G.; Scherer, S.; Ramanan, D.; and Luiten, J","cited_arxiv_id":null,"evidence_quote":"SplaTAM is the 3DGS SLAM baseline that fails on the full ScanNet++ S1 sequence, the key comparison for tracking robustness."},{"cited_title":"H.; and Davison, A","cited_arxiv_id":null,"evidence_quote":"MonoGS is the 3DGS pose-optimization baseline used in the convergence-basin comparison and a main tracking baseline on TUM."},{"cited_title":"T.; Mildenhall, B.; Verbin, D.; Srinivasan, P","cited_arxiv_id":null,"evidence_quote":"Origin of the depth-distortion term that the paper repurposes as the uncertainty measure for adaptive mapping."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GS-SLAM provides adaptive expansion and coarse-to-fine tracking ideas used in the surfel management design."}],"review_version":1}