{"id":"520f0653-2310-45d1-a4b7-dd26c7568340","arxiv_id":"2412.09868","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RP-SLAM reports state-of-the-art rendering quality and compact model size for 3DGS-based SLAM by combining gradient-guided adaptive sampling, KNN filtering, a dynamic keyframe window, and sparse-point-cloud monocular initialization.","lead":"RP-SLAM is a visual SLAM system that combines ORB-SLAM3 tracking with 3D Gaussian splatting mapping, adding adaptive gradient-guided sampling, KNN-based primitive filtering, a covisibility-based dynamic keyframe window, and sparse-point-cloud initialization for monocular cameras. The paper reports higher rendering quality, smaller map sizes, and real-time frame rates than five prior 3DGS SLAM systems on TUM, Replica, and ScanNet++.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Monocular bootstrap depends on sparse-point depth initialization with no specified handling of undefined depth; robustness on feature-poor or textureless regions is unestablished.","rationale":"The paper's central claim is that RP-SLAM achieves state-of-the-art rendering accuracy with real-time speed and compact models in both monocular and RGB-D modes. The RGB-D results are supported by strong numbers on Replica, TUM, and ScanNet++, but the monocular results are the more fragile part because they rely on a self-supervised bootstrap: sparse points initialize Gaussians, whose rendered depth then guides dense sampling. The reader's weakest assumption identifies exactly this dependency, and I agree that it is the most load-bearing concern. Nothing in the paper proves that the bootstrap succeeds when the sparse point cloud is thin or the scene has large textureless areas. The absence of error bars and the stochastic dynamic keyframe window further weaken the confidence in the reported margins, but the monocular bootstrap is a deeper issue because it threatens the method's validity, not just the precision of the numbers. The ablations show each module helps, yet they are on a single scene and do not isolate the failure mode of undefined depth. Therefore the appropriate verdict remains CONDITIONAL: the idea is plausible and the RGB-D evidence is encouraging, but the monocular claim needs a focused stress test and a specification of the undefined-depth policy. My read does not change the reader's verdict; it sharpens the condition under which acceptance would be justified.","tokens_in":19716,"tokens_out":7192,"duration_ms":80090,"concrete_test":"Run RP-SLAM on a monocular TUM sequence with large textureless regions (e.g., fr3_nostructure_texture_near_withloop) and on Replica with low-gradient regions artificially masked, comparing PSNR against Photo-SLAM and CaRtGS. Also instrument the pipeline to report the fraction of sampled pixels with no rendered depth under the current implementation and test two fallback policies (skip vs. nearest-depth). If PSNR drops by more than 1 dB relative to baselines, or if the undefined-depth fraction exceeds 20% on a standard sequence, the monocular bootstrap is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section III-D, the monocular keyframe initialization places Gaussians from ORB-SLAM3 sparse points, renders an initial depth map via Eq. 2, and then uses this depth to back-project adaptively sampled pixels (Eq. 6). However, a sparse set of Gaussians yields undefined (or zero) depth for most pixels in textureless or unobserved regions. The paper never specifies how such pixels are treated during back-projection or KNN filtering. If undefined depths are skipped, texture-rich regions without initial points are never filled; if arbitrary depths are assigned, the resulting Gaussians are misplaced and errors compound through subsequent optimization. The claimed monocular SOTA (PSNR 35.31 on Replica, 21.89 on TUM) rests on this bootstrap being reliable, yet the paper offers only one qualitative example (Fig. 3) and no sensitivity analysis to sparse-point density or accuracy. Since MKI, EIM, and DKW all interact through this depth bootstrap in monocular mode, a failure here would directly undermine the monocular half of the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents RP-SLAM, a decoupled 3D Gaussian Splatting SLAM system for monocular and RGB-D cameras. Camera poses are obtained from ORB-SLAM3, while the map is represented by Gaussian primitives that are optimized with three proposed components: an efficient incremental mapping module (quadtree-based adaptive sampling plus KNN-based Gaussian filtering), a dynamic keyframe window that randomly selects co-visible and non-co-visible keyframes to mitigate forgetting, and a monocular keyframe initialization that uses ORB-SLAM3 sparse points to bootstrap Gaussian placement and a rendered depth map. The authors evaluate on Replica, TUM, and ScanNet++ in both monocular and RGB-D settings, reporting state-of-the-art rendering accuracy (e.g., Replica RGB-D average PSNR 37.12 and LPIPS 0.056), compact model sizes around 10--11 MB, and real-time frame rates around 17--18 FPS. Ablation studies on Replica office0 support the contribution of each module, and a minimum-cell-size ablation shows the expected quality-efficiency trade-off.","tokens_in":19889,"tokens_out":4578,"duration_ms":52636,"significance":"If the claims are substantiated, RP-SLAM would be a useful step for real-time photorealistic mapping: it demonstrates that decoupling tracking from Gaussian optimization can yield compact maps with high rendering quality, and it offers a concrete monocular initialization mechanism that improves on random-depth initialization. The experimental breadth across three datasets and the use of official baseline implementations are strengths. However, the monocular bootstrap has an underspecified handling of undefined rendered depths, the reported averages have no error bars despite randomized keyframe selection, and the minimum cell size is tuned per dataset. These issues make the current evidence promising rather than definitive, and they are load-bearing for the central state-of-the-art claim.","major_comments":[{"comment":"The monocular keyframe initialization depends on rendering an initial depth map from a sparse set of Gaussians and then using that depth to back-project adaptively sampled pixels via Eq. (6). The paper never specifies how pixels for which Eq. (2) yields no depth are treated during back-projection or KNN filtering. If such pixels are skipped, texture-rich but initially unobserved regions are never filled; if arbitrary depths are assigned, misplaced Gaussians are optimized under Eq. (4) and the error compounds. This is the geometric foundation for the monocular results in Tables I and III, yet the only supporting evidence is the qualitative Fig. 3, with no sensitivity analysis to sparse-point density or accuracy. The authors should specify the exact depth-handling rule and provide a robustness study (e.g., degrading or densifying the ORB-SLAM3 point cloud) to support the monocular claim.","section":"Section III-D, Eq. (2), Eq. (6)"},{"comment":"The paper states that all results are averages of three runs, but no standard deviations or per-run values are reported. This matters because the dynamic keyframe window in Eq. (8) uses random selection, and several headline margins are small: for example, on TUM RGB-D the PSNR advantage over CaRtGS is 23.09 versus 22.11, and on Replica monocular the LPIPS difference is 0.067 versus 0.068. Without variance information, the reader cannot determine whether the reported improvements are statistically meaningful. Adding error bars or per-sequence variance tables is necessary to support the state-of-the-art claim.","section":"Section IV-A, Tables I--IV"},{"comment":"The minimum cell size c is set to 8 on Replica and TUM but to 4 on ScanNet++, and Table VI shows that c strongly controls both PSNR (37.12 to 42.08) and model size (5.8 to 29.4 MB) on Replica office0. The paper gives this as a deliberate choice for high-resolution details but does not provide a selection rule or show that the ScanNet++ comparison is robust to this choice. Since the ScanNet++ results in Table IV are a major part of the claimed advantage, the authors should either report results under the same default hyperparameter or justify the per-dataset choice with a principled criterion rather than post hoc tuning.","section":"Section IV-A (Implementation Details), Table VI"},{"comment":"The module ablation is performed on a single scene (Replica office0) in monocular mode only. This is too limited to establish that the proposed modules generalize across sequences, sensor modes, and scene types, especially because the EIM-only configuration reduces PSNR relative to the all-disabled baseline (34.23 versus 34.87), indicating that EIM is primarily an efficiency mechanism. The paper should include ablations on at least one additional Replica scene and one real-world sequence, and should report SSIM, LPIPS, ATE, and FPS in the ablation to verify that the quality improvements are not confined to PSNR on one scene.","section":"Section IV-E, Table V"}],"minor_comments":[{"comment":"The text uses \"3D Gaussian Sputtering\" instead of \"3D Gaussian Splatting\"; please correct this typo.","section":"Section I"},{"comment":"\"Datesets\" should be \"Datasets\" in the experimental setup heading.","section":"Section IV-A"},{"comment":"The RGB-D loss term contains the typo \"Lgoe\"; this should read \"Lgeo\".","section":"Eq. (4)"},{"comment":"The notation for the quadtree stopping criteria is confusing: the text defines \"adaptive minimum cell size cth = ηc\" and \"adaptive threshold τth = ητ\", but the subdivision condition is then stated as a variance comparison V(C) > τ. Please clarify how cth and τth are used, and whether c itself is a minimum cell size in pixels or a dimensionless parameter.","section":"Section III-B"},{"comment":"The ATE units are stated in the text as centimeters, but the tables do not repeat the units; please add the unit to the table headers or captions.","section":"Tables I--IV"},{"comment":"The paper does not mention a code release. Given that the implementation builds on the MonoGS codebase and several hyperparameters (τ, λ, k1, k2, c) are introduced, a code release or detailed implementation note would substantially improve reproducibility.","section":"Authorship and reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the experimental breadth is a strength, but the monocular initialization description has a concrete gap (undefined depths in the back-projection step) and the statistical basis for the state-of-the-art claim is weak without error bars. I would encourage the editor to request the authors to address the undefined-depth handling explicitly and to add repeated-run statistics, as these are fixable within the manuscript's scope. Code release would also materially increase confidence in the results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent engineering contribution to 3DGS-based SLAM. The three proposed components—gradient-guided quadtree sampling, KNN primitive filtering, and a covisibility-based dynamic keyframe window—are individually simple, but they fit together well. The monocular initialization from ORB-SLAM3 sparse points is the most interesting piece, and the ablations suggest it adds real value over random-depth baselines.\n\nWhat's solid: the experimental coverage is decent. They compare against five methods using official code, on three datasets, and report PSNR/SSIM/LPIPS plus model size and FPS. The model size reductions are substantial (e.g., ~10 MB vs 20-30 MB on Replica) without a clear rendering-quality trade-off. The ScanNet++ numbers are particularly strong. The ablation in Table V supports each module, though on a single scene. The citation pattern is clean; the one self-citation is unrelated to the central claims.\n\nSoft spots: no code release, no error bars even though they report three-run averages, and the quadtree minimum cell size is tuned per dataset (c=8 for Replica/TUM, c=4 for ScanNet++), which weakens the generality claim. The stress-test concern about monocular initialization is fair: the paper doesn't specify what happens when a sampled pixel has no valid rendered depth after the sparse-point bootstrap. In texture-poor regions where ORB gives few features, this could leave holes or create spurious Gaussians. That said, adaptive sampling deliberately avoids low-gradient areas, so the practical impact may be small. It's a missing detail, not a demonstrated flaw.\n\nThe \"state-of-the-art\" claim is a bit stronger than the evidence on TUM, where RP-SLAM wins PSNR but not SSIM/LPIPS. On Replica and ScanNet++ it's more defensible.\n\nBottom line: this deserves a serious referee. It's a useful, reproducible-in-principle method and the central claims are plausible. I'd request code and robustness analysis in revision, and my own verdict would be conditional acceptance.","headline":"Useful 3DGS-SLAM contribution with a genuinely interesting monocular initialization scheme; the SOTA claim is mostly credible but needs code, error bars, and a few robustness details.","tokens_in":20481,"tokens_out":3574,"would_cite":true,"duration_ms":32997,"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":"RP-SLAM reports state-of-the-art photorealistic mapping in real time by separating camera tracking from Gaussian map optimization, initializing monocular maps from sparse points, and pruning redundant Gaussians.","keywords":["3D Gaussian Splatting","visual SLAM","photorealistic mapping","monocular SLAM","RGB-D SLAM","incremental mapping","keyframe selection","real-time rendering"],"falsifier":"Run the monocular pipeline on a low-texture indoor corridor where the feature tracker produces only a handful of points per keyframe; if the rendered depth used for initialization is visibly empty or wrong and final PSNR falls far below the RGB-D result on the same scene, the sparse-point-cloud initialization is the step that fails.","tokens_in":19494,"feed_emoji":"🗺️","tokens_out":7949,"duration_ms":75414,"temperature":0.7,"pith_summary":"RP-SLAM is a visual SLAM system that builds a photorealistic 3D map with 3D Gaussian Splatting while running fast enough for real-time use and keeping the stored map small. The paper's central claim is that by decoupling camera-pose tracking from Gaussian optimization, and by sampling image pixels adaptively rather than densely, a 3DGS-based map can be both more accurate and far more compact than those of existing coupled or fixed-window systems. It reports the best average rendering quality on the Replica RGB-D benchmark (37.12 dB PSNR, 0.056 LPIPS) at about 18 frames per second and a model size near 10 MB, and it shows the same advantages in monocular mode, where a new initialization from the tracker's sparse point cloud replaces random-depth initialization. If correct, the result means photorealistic dense mapping is within reach of practical, real-time robot and AR/VR systems rather than requiring offline or heavily coupled pipelines.","feed_headline":"Photorealistic SLAM shrinks to 10 MB while running at 18 FPS","feed_subtitle":"Adaptive sampling and dynamic keyframe windows keep 3D Gaussian maps compact, accurate, and forgetting-free.","key_machinery":"The load-bearing object is the set of 3D Gaussian primitives, each an ellipsoid with a position, covariance (scale and quaternion), color, and opacity, rasterized by $\\alpha$-blending. The argument is carried by three mechanisms attached to that object: quadtree-based adaptive sampling, where each keyframe is subdivided until a gradient-variance threshold or a minimum cell size is reached and sampled pixels are back-projected into candidate primitives; KNN-based primitive filtering, which discards a new primitive when it lies within a scaled radius $\\lambda r_i$ of all $k=3$ nearest existing primitives; and a dynamic keyframe window, which each iteration forms $\\mathcal{W} = \\{K_{new}\\} \\cup S_1 \\cup S_2$ from $k_1$ co-visible and $k_2$ non-co-visible keyframes so optimization revisits old regions. In monocular mode, a fourth mechanism uses the tracker's sparse point cloud to place initial primitives and render a provisional depth via $\\alpha$-blended rasterization, which then feeds the adaptive sampling.","core_discovery":"On its own terms, the discovery is that the three failure modes of 3DGS-based SLAM—redundant primitives, forgetting during continuous optimization, and lack of geometric depth for monocular initialization—can be handled without sacrificing real-time speed. The paper proposes an incremental mapping stage that samples pixels with a quadtree guided by local gradient variance, so textured regions get more samples, and then filters new Gaussian primitives by a KNN test that discards any candidate lying within a scaled radius of all its three nearest neighbors. It replaces the fixed keyframe window with a dynamic window that randomly draws a fixed number of co-visible and non-co-visible keyframes each iteration, preserving both local and global consistency. For monocular cameras, it initializes Gaussians from the sparse point cloud of the feature tracker, renders an initial depth map, and uses that depth to densify the new region. The reported experiments claim state-of-the-art rendering accuracy on Replica, TUM, and ScanNet++ in both RGB-D and monocular modes, with model sizes several times smaller than most baselines.","pith_inferences":["A consequence the authors leave implicit is that the same quadtree sampling plus KNN filter could be dropped into any 3DGS mapping front-end, not only SLAM, wherever incremental dense reconstruction is needed.","The monocular initialization could be stress-tested by combining it with a learned dense depth prior; the paper only uses the sparse tracker points, so scenes with few features are the natural failure boundary.","The dynamic window could be made deterministic by selecting keyframes with maximal information gain or forgetting risk rather than random draws, which would make the consistency gain more reproducible on a single run.","A testable extension would be to measure forgetting directly: render revisited regions at the end of a long trajectory and compare PSNR to their first-render values, since the paper reports only aggregate quality."],"forward_implications":["On Replica RGB-D, the system reports the highest average PSNR (37.12) and lowest LPIPS (0.056) among compared methods, at 18.3 FPS with a 10.0 MB model.","In monocular mode on Replica, it reports 35.31 dB PSNR, substantially above the compared coupled and densification-based monocular systems, while keeping model size near 11 MB.","On ScanNet++, the highest PSNR is reported for both training and novel views, with roughly a quarter of the model size of the next-best compared method.","The full ablation on Replica office 0 shows the three modules are complementary: together they raise PSNR from 34.87 to 37.74 dB and cut model size from 19.3 to 11.8 MB."],"supporting_citations":[{"why":"Supplies the feature-based tracker that provides camera poses and the sparse point cloud used for monocular keyframe initialization.","marker":"[1]"},{"why":"Provides the 3D Gaussian primitive representation and differentiable tile rasterization on which the whole map is built.","marker":"[12]"},{"why":"Serves as an RGB-D baseline with a much larger model size, used to demonstrate the compactness advantage.","marker":"[13]"},{"why":"Supplies the base 3DGS-SLAM pipeline and hyperparameters whose modules the ablation study replaces one at a time.","marker":"[14]"},{"why":"Serves as a fast RGB-D baseline with high SSIM, compared on model size, frame rate, and rendering quality.","marker":"[15]"},{"why":"One of the decoupled baselines whose gradient-densification monocular initialization is the approach the paper argues against.","marker":"[16]"},{"why":"A decoupled baseline with strong monocular results; the paper compares model size and rendering accuracy against it.","marker":"[20]"},{"why":"Supplies the real-world low-quality RGB-D sequences used to test robustness under motion blur and sparse depth.","marker":"[21]"},{"why":"Supplies the synthetic Replica benchmark used for the main monocular and RGB-D quantitative comparisons.","marker":"[22]"},{"why":"Supplies the high-resolution ScanNet++ benchmark used to show rendering quality on real indoor scenes and novel views.","marker":"[23]"}],"fun_headline_variants":["Real-time photorealistic SLAM with leaner Gaussian maps","RP-SLAM: Photorealistic mapping without forgetting, at 18 FPS","Tiny 3D Gaussian maps keep SLAM photorealistic and real-time","Monocular and RGB-D SLAM gets compact, accurate Gaussian maps","SLAM that prunes Gaussians, avoids forgetting, and runs real-time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the sparse point cloud from the feature tracker is dense and accurate enough, especially in monocular mode, that the depth image rendered from it is a reliable geometric foundation for initializing new Gaussians; where features are scarce or scale is wrong, this foundation weakens.","fun_headline_variants_meta":{"raw":{"variants":["Real-time photorealistic SLAM with leaner Gaussian maps","RP-SLAM: Photorealistic mapping without forgetting, at 18 FPS","Tiny 3D Gaussian maps keep SLAM photorealistic and real-time","Monocular and RGB-D SLAM gets compact, accurate Gaussian maps","SLAM that prunes Gaussians, avoids forgetting, and runs real-time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1719,"prompt_tokens":1002,"completion_tokens":717,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":618}},"tokens_in":618,"tokens_out":717,"duration_ms":7914,"temperature":1.0,"reasoning_tokens":618,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:37:32.412883+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the monocular pipeline on a low-texture indoor corridor where the feature tracker produces only a handful of points per keyframe; if the rendered depth used for initialization is visibly empty or wrong and final PSNR falls far below the RGB-D result on the same scene, the sparse-point-cloud initialization is the step that fails.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering,","cited_arxiv_id":null,"evidence_quote":"Provides the 3D Gaussian primitive representation and differentiable tile rasterization on which the whole map is built."},{"cited_title":"Rtg- slam: Real-time 3d reconstruction at scale using gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Serves as a fast RGB-D baseline with high SSIM, compared on model size, frame rate, and rendering quality."},{"cited_title":"Photo-slam: Real- time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras,","cited_arxiv_id":null,"evidence_quote":"One of the decoupled baselines whose gradient-densification monocular initialization is the approach the paper argues against."},{"cited_title":"Scannet++: A high- fidelity dataset of 3d indoor scenes,","cited_arxiv_id":null,"evidence_quote":"Supplies the high-resolution ScanNet++ benchmark used to show rendering quality on real indoor scenes and novel views."}],"review_version":1}