{"id":"2acdef1d-dd3d-4476-be2f-bc2595e4fa22","arxiv_id":"2511.23221","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CB-KNN smooths the colors and positions of nearby Gaussians during keyframe rendering and reports modestly lower tracking error on Replica, TUM-RGBD, and ScanNet.","lead":"This paper adds a smoothing step to a 3D Gaussian Splatting SLAM system: before rendering keyframes, it averages the colors and nudge positions of the K nearest Gaussians so pose tracking is less sensitive to bad map parameters. The authors report lower trajectory errors on three RGB-D benchmarks with roughly unchanged reconstruction quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates CB-KNN: ATE gains vs SplaTAM could stem from changed keyframe selection or loss weighting, not the smoothing.","rationale":"The reader's weakest_assumption focuses on the undefined 2D-to-3D lift and gradient flow in Eq. (6). That is a real methodological gap that prevents reproduction and casts doubt on whether the implementation is mathematically sound. However, I judge the lack of a controlled ablation to be the more directly load-bearing concern for the central claim. Even if the correction were fully specified, the experimental evidence would still be confounded by simultaneous changes to keyframe selection and the loss function. The paper presents no run 'with other changes but without CB-KNN,' so the ATE improvements could be entirely due to these other modifications. This is a standard and severe threat to internal validity. I therefore agree partially with the reader: the implementation gap is important, but the ablation gap is the one most likely to overturn the central claim. The concrete test proposed would settle whether the smoothing step itself carries the benefit, which is essential before accepting the paper's conceptual conclusion. Since the reader already conditioned acceptance on ablations and code, my concern supports the existing CONDITIONAL verdict rather than changing it.","tokens_in":13967,"tokens_out":9878,"duration_ms":97522,"concrete_test":"Run the authors' pipeline on Replica R0/R1 and TUM-RGBD fr1/desk with CB-KNN disabled (set α=0 or skip the position/color correction) while retaining every-nth-frame keyframe selection and the 0.6 color weight. Compare the resulting ATE RMSE to (a) original SplaTAM and (b) the full CB-KNN pipeline. If the disabled-CB-KNN run matches the full pipeline's ATE, the improvement is not due to CB-KNN; if it matches SplaTAM, the smoothing is the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that CB-KNN smoothing during keyframe rendering is responsible for the ATE reductions reported in Table I (e.g., 0.39→0.32 cm on Replica). But the comparison to SplaTAM is not controlled: the authors also changed keyframe selection (to every nth frame, Section III-C) and added a color loss weight of 0.6 (Eq. 8, tuned via repeated experiments). SplaTAM uses its own overlap-based keyframe selection and a different loss. Without an ablation that disables CB-KNN while keeping these other changes, the observed improvements cannot be uniquely attributed to the smoothing step. This is the load-bearing gap: if the α=0 (no-correction) pipeline with the same keyframe/loss settings performs as well as the full CB-KNN, the paper's core hypothesis is not supported by the evidence. A related side inconsistency appears in Eq. (10): the text says K should increase when Gaussian density is low or scene complexity is high, but the formula K=K0·max(0.5, 1−βγ/(ρ+ε)) decreases K exactly in those regimes (since γ/(ρ+ε) grows), which suggests the implementation may not match the stated adaptive behavior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper challenges the assumption that rendering fidelity is the main driver of tracking accuracy in 3DGS-based SLAM. It proposes Corrective Blurry KNN (CB-KNN), a temporary correction applied during keyframe rendering that shifts the 2D projections of the K-nearest Gaussians toward their centroid and replaces their colors with a contribution-weighted average. The authors claim this controlled blur/densification makes pose optimization more robust to Gaussian parameter noise without modifying the underlying map. They report ATE RMSE reductions versus SplaTAM on Replica (0.39→0.32 cm), TUM-RGBD (3.31→3.18 cm), and ScanNet (13.43→12.48 cm), comparable or slightly improved rendering metrics, and lower runtime on Replica/R0. The paper includes an adaptive K formula depending on local Gaussian density and inter-frame motion.","tokens_in":14304,"tokens_out":4872,"duration_ms":48106,"significance":"If the central claim is upheld, the paper offers a simple, modular mechanism that could be plugged into existing 3DGS-SLAM systems to improve tracking robustness at negligible mapping cost. The empirical direction is consistent across three benchmarks, which is encouraging. The paper also provides constructive specificity by naming its hyperparameters (K0, β, α, color weight) and by restricting the correction to keyframes, which makes the proposal concrete and testable. However, the current evidence does not isolate the proposed mechanism: the comparison against SplaTAM involves several simultaneous changes, key geometric/differentiability details of the correction are undefined, and no error bars or ablations are provided. The significance of the contribution is therefore real but conditional on resolving these load-bearing issues.","major_comments":[{"comment":"Equation (6) defines the corrected position only through its 2D projection: π(µ'_pk)=π(µ_pk)+α(...). No 3D vector µ'_pk is ever specified, yet Eqs. (2)–(4) use f_pk(p) for the \"corrected Gaussian\" and the text says f_pk uses the corrected 2D projection position as input. Since Eq. (1) defines the Gaussian decay in 3D, and the rasterizer in 3DGS requires a 3D mean (or a well-defined 2D Gaussian with covariance), the rendering operation is underspecified. More critically, the paper claims pose and map optimization remain end-to-end differentiable through the corrected rendering. No Jacobian of the correction with respect to µ_pk or E_t is given. Please define the 3D update (e.g., depth-based unprojection of the 2D offset), the covariance used after correction, and the gradient derivation. Without this, the reported ATE gains may be artifacts of an unspecified implementation.","section":"§III-B, Eq. (6)"},{"comment":"The adaptive-K description and formula are contradictory. The text states that K should increase when Gaussian density is low or scene complexity is high (\"increasing the K value enhances the smoothing effect when Gaussian density is low or scene complexity is high\"); however K=K0·max(0.5, 1−βγ/(ρ+ε)) decreases as γ/(ρ+ε) grows, i.e., in exactly those regimes. Either the formula or the stated intuition is wrong, and this matters because the claimed adaptivity is a named contribution. Please align the two and report which behavior was actually implemented in the experiments.","section":"§III-C, Eq. (10)"},{"comment":"The central claim is that CB-KNN smoothing is responsible for the ATE reductions, but the comparison to SplaTAM is uncontrolled. The pipeline also changes keyframe selection (every nth frame vs. SplaTAM's overlap-based), the loss (color weight 0.6 in Eq. (8), tuned by repeated experiments), keyframe count in map updates, and K0 per dataset. No ablation disables CB-KNN (e.g., α=0 or no position/color correction) while keeping the other modifications. Without that ablation, the improvement cannot be attributed to the smoothing step; it could come from any of the other differences. This experiment is essential for the paper's hypothesis.","section":"§IV, Table I; §III-C"},{"comment":"Experimental support is statistically thin. All metrics are single runs without error bars, and several hyperparameters (K0=5 vs. 8, color weight 0.6, β=0.3) are tuned on the same benchmark suites. The reported gains are small (e.g., Replica ATE 0.39→0.32 cm, TUM 3.31→3.18 cm), so without multiple runs or a validation protocol one cannot exclude that the differences are due to noise or to tuning. Please report mean±std over at least 3 runs and separate validation/test selection of hyperparameters.","section":"§IV-A, Tables I–III"}],"minor_comments":[{"comment":"Duplicate phrase in \"utilizing rapidly rendered rapidly rendered color maps and depth maps.\"","section":"§II-B"},{"comment":"The acronym \"CK-KNN\" is used in the Relative Pose Error Analysis section; it should be CB-KNN.","section":"§IV-B"},{"comment":"The loss expression appears to be missing a plus sign and a parenthesis: it should read L1(D(p)−D_GT(p)) + 0.6·L1(C(p)−C_GT(p)), not L1(D(p)−D_GT(p) + 0.6·L1(C(p)−C_GT(p))).","section":"Eq. (8)"},{"comment":"K0 is inconsistently specified: §III-C says the baseline value K0=8, while §IV-A states K0=5 for Replica and K0=8 for TUM-RGBD/ScanNet. Please clarify which value is used and why.","section":"§III-C and §IV-A"},{"comment":"Reference [15] and [25] are the same paper (Woodford & Rosten); Fig. 3's caption cites SplaTAM as [9] instead of [10]; some table entries are typeset inconsistently (e.g., lowercase scene names in Table III).","section":"References and captions"}],"recommendation":"major_revision","confidential_remarks":"The main risk with this manuscript is attribution: the core empirical claim—that CB-KNN is the cause of the improved ATE—is not tested in isolation, and the geometric/differentiability details of Eq. (6) are not sufficient to reimplement the method. If the authors provide the requested ablation, the full 3D correction and gradient definitions, and a small statistical evaluation with error bars, the contribution could become a useful practical addition to 3DGS-SLAM. In its current form, the paper is better suited as a workshop preprint than a journal publication. I encourage the editor to require the missing ablation before considering it further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The CB-KNN idea is neat, but this paper doesn't actually show that CB-KNN is doing the work. Worth a serious referee, though it needs major revision.\n\nWhat's new: the authors argue that for pose tracking in 3DGS-SLAM, a slightly blurred rendering can be more robust than a sharp one, and propose a transient K-nearest-Gaussian smoothing of positions and colors during keyframe rendering. That's a different objective from the usual rendering-quality focus, and the method integrates cleanly with existing splatting pipelines. On Replica, TUM-RGBD, and ScanNet they report consistent average ATE improvements over SplaTAM (0.39 to 0.32 cm, 3.31 to 3.18 cm, 13.43 to 12.48 cm) with similar or slightly better rendering metrics. Good.\n\nBut the experiments don't isolate the smoothing. Compared with SplaTAM they also switched to every-nth-frame keyframe selection and added a color loss weight of 0.6, tuned via repeated experiments. Without an ablation that turns CB-KNN off (alpha=0) while keeping those settings, the ATE gains could come from either change or from hyperparameter tuning on the same benchmarks. The gains are small and there are no error bars or multi-run statistics. That is the load-bearing gap.\n\nThere's also a real internal inconsistency in Eq. (10). The text says K should increase when Gaussian density is low or scene complexity is high, but the formula K = K0 * max(0.5, 1 - beta*gamma/(rho+epsilon)) decreases K exactly in those regimes. Either the formula or the description is wrong, and the authors need to reconcile that.\n\nFinally, Eq. (6) defines the corrected Gaussian position only through its 2D projection; how that maps back to the 3D position used in attenuation and depth terms, and how gradients flow w.r.t. pose and map parameters, is never made explicit. So the differentiability claim is unverified.\n\nNone of this kills the core idea — smoothing during rendering as regularization for tracking is plausible. But the current paper overclaims \"significantly improves\" without controlled evidence. I'd send it to peer review, but the revision needs an ablation, code, and a fix to Eq. (10). Researchers in dense RGB-D or 3DGS-SLAM would be the audience; they should read it as a promising direction rather than a validated method.","headline":"Plausible idea, uncontrolled evidence: CB-KNN smoothing could help 3DGS-SLAM tracking, but no ablation isolates it, and Eq. (10) contradicts its own adaptive story.","tokens_in":14747,"tokens_out":3422,"would_cite":false,"duration_ms":31233,"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 transient, local blur of rendered keyframes makes 3D Gaussian Splatting SLAM track more accurately, without touching the underlying map.","keywords":["3D Gaussian Splatting","SLAM","camera pose tracking","adaptive kernel smoothing","K-nearest neighbor","rendering robustness","RGB-D SLAM","keyframe optimization"],"falsifier":"Render a keyframe with CB-KNN enabled but detach the gradients of the corrected positions and colors (stop-gradient), so the optimizer sees the smoothed image but receives no gradient from the correction. If pose tracking accuracy remains the same, the smoothing effect alone—not its differentiability—is what helps; if accuracy degrades, the gradient through the correction is load-bearing. Alternatively, compute the Jacobian of the rendered color with respect to the original Gaussian center through Eq. (6): if that Jacobian is inconsistent with the actual 3D position used in the attenuation fun","tokens_in":13838,"feed_emoji":"🎯","tokens_out":4673,"duration_ms":40884,"temperature":0.7,"pith_summary":"This paper challenges the assumption that higher rendering fidelity automatically yields better pose tracking in 3DGS-based SLAM. It argues that making the rasterization process robust to Gaussian parameter errors is more critical than perfect scene representation, and that a controlled, local smoothing of rendered keyframes can stabilize pose optimization. The proposed CB-KNN method adaptively shifts the 2D projections of the K-nearest Gaussians toward their centroid and blends their colors, applied only during keyframe rendering. On Replica, TUM-RGBD, and ScanNet, this reduces trajectory error while preserving or slightly improving rendering quality, supporting the claim that temporary rendering regularization is a practical alternative to redesigning the rasterizer.","feed_headline":"Smoothing keyframes cuts pose error in 3DGS-SLAM","feed_subtitle":"A transient K-nearest-neighbor blur during rendering lowers trajectory error on Replica, TUM-RGBD, and ScanNet without touching the map.","key_machinery":"CB-KNN (Corrective Blurry K-Nearest Neighbors): a per-pixel correction applied at render time. Select the K Gaussians with highest contribution, compute the centroid of their 2D projections, shift each projection by a fraction α of its distance to the centroid (Eq. 6), and set each color to the normalized attenuation-weighted average of the neighbors (Eq. 7). This produces a smoothed temporary Gaussian set used only for that frame's rendering, preserving differentiability for pose and map optimization while not altering the stored map.","core_discovery":"The paper establishes that a 'Corrective Blurry KNN' (CB-KNN) pre-rendering step—which, for each pixel, nudges the K nearest Gaussians toward their common centroid in the image plane and replaces their colors with a contribution-weighted average—acts as a regularizer that dampens parameter noise from outlier Gaussians. This yields smoother depth/color/contour images for pose optimization, improving ATE RMSE from 0.39 to 0.32 cm on Replica, 3.31 to 3.18 cm on TUM-RGBD, and 13.43 to 12.48 cm on ScanNet, with comparable or better rendering metrics. The corrections are transient and leave the map parameters unchanged.","pith_inferences":["The paper leaves unspecified how the corrected 2D projection is lifted back to a 3D position for the attenuation function; a careful reader should check whether the implementation uses an approximate depth or a projective lifting, and whether the gradient path is consistent with the stated formulas.","A natural extension is to test CB-KNN in stereo or monocular 3DGS-SLAM, where depth is estimated and parameter noise is higher; the smoothing may yield larger gains there.","The robustness mechanism suggests an analogy to robust estimation: CB-KNN acts as a nonparametric outlier filter on Gaussian parameters; one could try other robust statistics (median, trimmed mean) in the same slot.","The adaptive K formula depends on two hand-set constants (K0, β); ablating them could reveal whether the adaptive selection is the active ingredient or simply the fixed blur level."],"forward_implications":["If the central claim is right, 3DGS-SLAM systems can be made more robust without modifying the map or the rasterizer core, just by inserting a lightweight correction at keyframe rendering.","The separation of mapping quality and tracking quality suggests that future SLAM systems may explicitly optimize for trackability, e.g., rendering with regularization rather than only for photometric accuracy.","CB-KNN's adaptive K selection based on local density and motion amplitude indicates the smoothing strength can be tuned per-frame, potentially generalizing to other 3DGS applications where robustness matters.","Improved LPIPS on Replica and PSNR on TUM suggest that smoothing during keyframe optimization can also improve the learned map's rendering quality, implying a beneficial side effect.","Because the correction is transient and keyframe-only, the added computational cost is small; the paper reports higher FPS and lower per-frame time than SplaTAM on Replica/R0."],"fun_headline_variants":["Blurring K-nearest Gaussians stabilizes 3DGS pose tracking","Pre-render Gaussian blur cuts track error in 3DGS-SLAM","Adaptive kernel smooths Gaussians for robust SLAM tracking","Temporary Gaussian blur lowers camera pose error in 3DGS-SLAM","Blend nearest Gaussians to steady 3DGS camera tracking"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The corrected Gaussian position is specified only in terms of its 2D projection, so the paper never defines how that offset is turned back into a 3D location for the attenuation function, nor how gradients flow through that corrected projection during pose and map optimization. If that step is not implemented with a valid differentiable 3D correction, the reported improvements could stem from the implementation choice rather than from the smoothing principle.","fun_headline_variants_meta":{"raw":{"variants":["Blurring K-nearest Gaussians stabilizes 3DGS pose tracking","Pre-render Gaussian blur cuts track error in 3DGS-SLAM","Adaptive kernel smooths Gaussians for robust SLAM tracking","Temporary Gaussian blur lowers camera pose error in 3DGS-SLAM","Blend nearest Gaussians to steady 3DGS camera tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000447,"raw_usage":{"total_tokens":2137,"prompt_tokens":831,"completion_tokens":1306,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":1207}},"tokens_in":575,"tokens_out":1306,"duration_ms":9213,"temperature":1.0,"reasoning_tokens":1207,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:33:22.379842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a keyframe with CB-KNN enabled but detach the gradients of the corrected positions and colors (stop-gradient), so the optimizer sees the smoothed image but receives no gradient from the correction. If pose tracking accuracy remains the same, the smoothing effect alone—not its differentiability—is what helps; if accuracy degrades, the gradient through the correction is load-bearing. Alternatively, compute the Jacobian of the rendered color with respect to the original Gaussian center through Eq. (6): if that Jacobian is inconsistent with the actual 3D position used in the attenuation fun","supporting_citations":[],"review_version":1}