{"id":"e92a1f00-a932-4f88-94ac-223148ec2155","arxiv_id":"2411.12185","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"LiV-GS couples LiDAR depth with camera color in a 3D Gaussian map and performs outdoor SLAM plus rendering at about 7.98 frames per second.","lead":"LiV-GS is a robot mapping system that combines a laser scanner and a camera to build a 3D map of outdoor scenes using colored Gaussian ellipsoids while simultaneously tracking the robot's path. It reports faster and more accurate mapping and rendering than several existing SLAM systems, and also demonstrates radar localization using the resulting map.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Out-of-FOV reconstruction rests entirely on Eq. (11), which is stated as a postulate with no specified update rule and no metric restricted to unobserved regions; the claimed advantage over prior LiDAR-visual 3DGS SLAM is therefore unvalidated.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing point: the Conditional Gaussian Constraint in Eq. (11) is the mechanism behind the paper's distinguishing claim of reconstructing regions beyond the LiDAR field of view, and it is the least supported component of the system. My stress-test pass confirms this and sharpens it: Eq. (11) is presented as a postulate, not a derived update rule; the paper does not explain how the constraint is incorporated into the optimization loss in Eq. (10); and the reliability propagation rule, namely that a split Gaussian becomes reliable after one round of backend optimization, has no geometric justification. These are internal incompletenesses in the argument, not merely disagreements with prior work. Because the claimed advantage over the closest baselines in Table I hinges on this mechanism, the lack of a targeted evaluation is a real soft spot. The concern does not overturn the paper's plausibility: the system may work, and the reported tracking and rendering numbers are competitive. However, the evidence as written does not establish the out-of-FOV reconstruction claim at full strength. Since the reader already assigned CONDITIONAL for essentially this reason, my verdict remains UNCHANGED rather than moving to a stronger rejection: the identified issue is an evidence gap that a focused experiment could close, not a demonstrated contradiction. Other weaknesses noted by the reader, such as missing error bars, segmented sequences, and absent comparisons to the closest recent methods, are secondary and would not by themselves change the verdict. The concrete test proposed above, a per-pixel in-FOV versus out-of-FOV evaluation plus a CGC ablation, directly targets the load-bearing assumption and would settle whether the central claim survives.","tokens_in":11772,"tokens_out":2651,"duration_ms":31646,"concrete_test":"Re-run LiV-GS on the cp and garden sequences while storing per-pixel LiDAR visibility masks for every keyframe. Compute rendering metrics (PSNR, SSIM, LPIPS) and rendered-depth RMSE separately for pixels with LiDAR depth and pixels without LiDAR depth, using held-out viewpoints whose LiDAR scans were not used in mapping. Also compute Chamfer distance or point-to-mesh error for out-of-FOV Gaussians against a later, registered dense LiDAR pass. Repeat with the Conditional Gaussian Constraint disabled. If out-of-FOV metrics are not clearly better than rendering only within the LiDAR FOV, or if disabling the CGC does not degrade out-of-FOV geometry, the central advantage claimed in the abstract and Section I is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novel advantage over the closest related methods, as summarized in Table I, is reconstructing scenes beyond the LiDAR field of view. This advantage depends entirely on the Conditional Gaussian Constraint in Section III-D, Eq. (11), which postulates p(X|Y=y) ~ N(mu_x(y), Sigma_y) for a color-supervised Gaussian x given its nearest LiDAR-measured Gaussian y. The paper never derives this conditional distribution, never states how it enters the loss function in Eq. (10), and never defines the reliability criterion beyond asserting that a split Gaussian becomes reliable after one round of backend optimization. In outdoor scenes with depth discontinuities, such as building edges, foliage, poles, and object silhouettes, a single nearest-neighbor covariance cannot encode whether the unobserved surface is coplanar with, adjacent to, or detached from the observed surface. Color continuity alone can produce visually plausible renderings while the geometry in out-of-FOV regions is wrong. The only geometric validation, the radar relocalization experiment in Section IV-F, is qualitative and tests the LiDAR-supported portion of the map rather than the extrapolated regions. No ablation disables the CGC, and no metric separates in-FOV from out-of-FOV pixels or rays. Thus the claim that LiV-GS reconstructs geometrically useful areas beyond the LiDAR field of view is supported only by the qualitative Fig. 4. This is an evidence gap rather than an internal contradiction, but it is load-bearing: if Eq. (11) merely hallucinates plausible texture into unobserved space, the paper's primary distinction from MM-Gaussian and related methods collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents LiV-GS, a LiDAR-visual SLAM system that uses 3D Gaussian splatting as the map representation for outdoor scenes. The front end tracks by aligning LiDAR point clouds with Gaussian centers using covariance and normal constraints; the back end optimizes poses and map parameters with photometric, geometric, and normal losses. A \"Conditional Gaussian Constraint\" (CGC) is introduced to propagate information from LiDAR-supported Gaussians to color-only Gaussians, aiming to reconstruct areas beyond the LiDAR field of view. Experiments on six sequences from NTU4DRadLM and one loop-closure sequence from R3LIVE report trajectory and rendering comparisons against several open-source baselines, plus a qualitative cross-modal radar relocalization result.","tokens_in":12086,"tokens_out":2192,"duration_ms":25037,"significance":"If the results hold, LiV-GS would be a useful contribution: it demonstrates that a 3DGS representation can support real-time LiDAR-visual SLAM outdoors, with rendering quality and tracking accuracy competitive with or better than several existing open-source systems. The reported 7.98 FPS processing rate is a concrete practical claim, and the cross-modal radar localization experiment is an interesting downstream use of a Gaussian map. However, the central claimed novelty—high-quality reconstruction beyond the LiDAR field of view—rests entirely on the CGC, whose validity is not quantitatively demonstrated. The significance of the paper therefore depends on closing that evidence gap and on comparing directly with the closest related LiDAR-visual 3DGS SLAM methods, which are only summarized in a table, not benchmarked.","major_comments":[{"comment":"The Conditional Gaussian Constraint is the load-bearing component for the paper's main claimed advantage over prior work (reconstruction beyond the LiDAR field of view, as stated in Table I and the abstract), but it is presented as a postulate with no derivation, no update rule, and no mechanism connecting it to the loss in Eq. (10). The paper states that a split Gaussian becomes reliable after one round of backend optimization, but this reliability criterion is not formally defined. The only evidence for the out-of-FOV reconstruction is qualitative Fig. 4. I request an ablation that removes the CGC, and a quantitative evaluation on pixels or rays whose LiDAR depth is unavailable, comparing in-FOV and out-of-FOV reconstruction quality. Without this, the claimed advantage is unsupported.","section":"III-D, Eq. (11)"},{"comment":"The closest related methods—DrivingGaussian, LIV-GaussMap, and MM-Gaussian, which are listed in Table I as the most relevant LiDAR-visual 3DGS fusion works—are not included in any quantitative comparison. Since these methods also fuse LiDAR point clouds with Gaussian maps and optimize poses, excluding them makes it impossible to verify the claim of superiority over the state of the art that this paper itself identifies. Please add quantitative trajectory and rendering comparisons with at least the open-source methods among them (e.g., MM-Gaussian), or clearly justify their exclusion if their public implementations are unavailable on the same platform.","section":"Table II and Table III"},{"comment":"The evaluation protocol has two issues that affect the reliability of the reported improvements. First, long sequences are segmented into shorter segments (e.g., 2100–2400 images) without reporting results on the full trajectories; this can favor methods that drift substantially over long distances, and it is unclear how the segment boundaries were chosen. Second, no error bars or multiple-run statistics are reported for any trajectory or rendering metric. Since several baseline methods are stochastic (e.g., MonoGS, SplaTAM) or initialization-sensitive, single-run comparisons are not sufficient to support the claim of superior or consistent performance. Please report mean/std over at least three runs, or state if the methods are deterministic and provide a rationale.","section":"IV-B and IV-C"},{"comment":"The runtime claim of 7.98 FPS is presented as evidence of real-time capability, but the system uses an asynchronous communication mechanism and the stated module times (0.07 ms, 0.04 ms, 0.09 ms) are implausibly small compared with the reported overall frame rate. It is unclear whether 7.98 FPS is measured under the asynchronous pipeline with all modules running concurrently, or as a throughput on a single thread. Please clarify the measurement protocol and report the actual wall-clock time per frame, as this claim is central to the abstract's \"fast and accurate\" characterization.","section":"IV-E and Fig. 7"}],"minor_comments":[{"comment":"The module name \"Data Preporessing\" in Fig. 1 is a typo; it should read \"Data Preprocessing.\"","section":"III-A"},{"comment":"The sentence \"integrates the normal orientation into the loss function to refines the Gaussian map\" has a grammar error; \"refines\" should be \"refine.\"","section":"Abstract"},{"comment":"The regularization term R is described as being multiplied by a hyperparameter, but no such hyperparameter appears in Eq. (3). Please define the full expression with its weight, and specify its value in the implementation details.","section":"III-B, Eq. (3)"},{"comment":"The term \"local spherical region Nr(x)\" is used in the density function definition, but the radius r is not specified. Please state how r is chosen or adapted.","section":"III-B"},{"comment":"Table II reports the rotational error in units of degrees per 100 m, but the units are not stated in the table caption. Please add the units to avoid ambiguity.","section":"IV-C"},{"comment":"The radar relocalization experiment is described as qualitative; Fig. 8 shows only the trajectory. Please provide quantitative relocalization error (e.g., ATE against the ground-truth trajectory) to substantiate the claim of accurate cross-modal localization.","section":"IV-F"},{"comment":"Some references are cited as arXiv preprints without version or DOI information (e.g., [4], [5], [10], [11], [14], [17], [21], [23], [29], [31]). This is acceptable for a preprint, but please update to published versions where available.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is timely, and the reported results are potentially interesting, but the main novelty (CGC) is the least validated part. The missing comparison with the closest related methods (MM-Gaussian, LIV-GaussMap, DrivingGaussian) is a significant gap for a robotics audience, and the segmented evaluation with no error bars weakens the empirical claims. The issue is fixable within the manuscript's scope: an ablation and an out-of-FOV metric would likely settle the concern. I would not recommend rejection if these validation points can be addressed. Also note that the 'first method' claim in the abstract is arguably overstated given the related work cited by the authors themselves; this should be softened or supported with a precise differentiation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: LiV-GS is a coherent system integration and the reported tracking/rendering numbers are competitive, but the one thing it claims over its closest relatives—reconstructing geometry beyond the LiDAR field of view—is not actually supported by the evidence in the paper. The Conditional Gaussian Constraint (Eq. 11) is a postulate, not a derived update rule, and there is no ablation or metric that isolates out-of-FOV regions. If that piece falls, the paper is a competent but largely incremental LiDAR-camera 3DGS SLAM.\n\nWhat is genuinely new: combining covariance-based point-to-Gaussian alignment (from GS-ICP-SLAM-ish work) with a normal-consistency weight, a normal-length regularizer, and conditional propagation of Gaussians into regions without LiDAR depth. The system runs at 7.98 FPS and produces the lowest ATE on several NTU4DRadLM sequences, and the rendering metrics beat the baselines tested. That is real, if modest, progress.\n\nSoft spots, in order of severity:\n- The central claim. Eq. (11) says given nearest LiDAR Gaussian y, color-supervised Gaussian x follows N(mu_x(y), Sigma_y). There is no derivation of this conditional, no statement of how it enters the loss, and no definition of 'reliable' beyond 'after one round of backend optimization.' At depth discontinuities—building edges, foliage, poles—a single nearest-neighbor covariance cannot encode whether the unobserved surface continues, is adjacent, or is detached. The only geometric validation, the radar relocalization, is qualitative and covers the LiDAR-supported portion of the map, not the extrapolated areas. This is exactly the load-bearing gap the stress-test note flags.\n- Evaluation gaps. No error bars. No comparison against MM-Gaussian, LIV-GaussMap, or Gaussian-LIC, the closest methods. Long sequences are segmented into 220–250 m chunks, which may help any frame-to-map tracker and avoids long-drift evaluation.\n- Minor: no code; the 7.98 FPS figure is a mean without variance; the skybox with 100,000 Gaussians is a detail that could hide cost.\n\nThe reader's concern about circularity is fair but not the main issue; this is a standard frame-to-map loop. The real problem is missing independent validation of the claimed advantage.\n\nWho this is for: anyone working on LiDAR-visual fusion for 3DGS SLAM. It deserves a serious referee—the system is plausible and the NTU numbers are worth checking—but the referee should require an ablation that disables the CGC and a separate metric for out-of-FOV regions, plus comparisons with the omitted related methods.","headline":"A plausible LiDAR-camera 3DGS SLAM system with competitive reported numbers, but its headline claim of reconstructing beyond the LiDAR FOV is not yet backed by evidence.","tokens_in":12681,"tokens_out":2019,"would_cite":true,"duration_ms":21221,"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":"LiV-GS proposes the first SLAM system that directly aligns sparse LiDAR point clouds with continuous 3D Gaussian maps in large outdoor scenes, and it reports 7.98 FPS mapping with rendering quality that beats the compared baselines.","keywords":["3D Gaussian Splatting","LiDAR-visual SLAM","outdoor mapping","novel view synthesis","conditional Gaussian constraint","Gaussian-LiDAR alignment","cross-modal localization","large-scale reconstruction"],"falsifier":"Measure the reconstructed surface in a region that LiDAR never observed and compare it against ground-truth geometry from a later dense LiDAR pass or a survey-grade scanner; if the conditional-Gaussian-split region deviates from the true surface by more than the in-field-of-view error, the central claim fails. A cheaper check is an ablation that removes the conditional Gaussian constraint and measures rendering quality on out-of-field-of-view image patches, since removing it should clearly degrade those regions if the constraint is doing the claimed work.","tokens_in":11514,"feed_emoji":"📡","tokens_out":6344,"duration_ms":59186,"temperature":0.7,"pith_summary":"This paper proposes LiV-GS, a simultaneous localization and mapping system that builds a 3D Gaussian Splatting map from LiDAR and camera data in outdoor scenes. The authors claim it is the first method to align sparse, discrete LiDAR point clouds directly with continuous differentiable Gaussian maps, avoiding the fixed-resolution limitation of conventional LiDAR maps. They further claim that their conditional Gaussian constraint lets the map grow into regions where LiDAR has no measurements, so the reconstruction extends beyond the LiDAR field of view. If these claims hold, a robot could localize and produce photorealistic, geometrically usable maps in near real time from a single pass of a LiDAR-camera pair.","feed_headline":"LiDAR-vision Gaussian map does outdoor SLAM at 7.98 FPS","feed_subtitle":"One continuous 3D Gaussian map handles tracking, mapping, and rendering, and reaches beyond the LiDAR field of view.","key_machinery":"The central object is the set of 3D Gaussian ellipsoids, each parameterized by opacity, color, mean, and covariance, used as the map and as the tracking target. Three mechanisms carry the argument: (1) covariance-based point-to-Gaussian matching with a density and normal-consistency weight W(x)=C(x)rho(x) and a normal-direction regularizer; (2) a normal loss that penalizes the variance along the Gaussian normal, keeping ellipsoids flat and surface-aligned; and (3) the Conditional Gaussian Constraint of Eq. (11), which propagates reliability from LiDAR-measured Gaussians to color-only Gaussians by sampling new ellipsoids from a normal distribution around the nearest reliable Gaussian. The constraint is the mechanism that extends reconstruction beyond the LiDAR field of view.","core_discovery":"LiV-GS treats a scene as a collection of 3D Gaussian ellipsoids and uses that same representation for both front-end tracking and back-end map optimization. LiDAR points are matched to Gaussian ellipsoids through shared covariance attributes, with a weighting function that combines local Gaussian density and normal consistency, plus a normal loss that keeps Gaussians flat along surfaces. The distinctive claim is the Conditional Gaussian Constraint: a color-supervised Gaussian outside the LiDAR field of view is pulled toward the nearest LiDAR-measured Gaussian by postulating p(X|Y=y) ~ N(mu_x(y), Sigma_y), and each split Gaussian is treated as reliable after one round of back-end optimization. This is what lets the map represent objects and regions never touched by LiDAR. The system reports 7.98 FPS average throughput and, on the tested sequences, generally lower trajectory error on low-speed sequences and better rendering metrics than the compared point-cloud, feature-based, and Gaussian Splatting SLAM baselines, and it demonstrates cross-modal radar localization on the resulting Gaussian map.","pith_inferences":["If the conditional Gaussian constraint propagation holds on longer and more varied sequences, it effectively turns visual texture into a geometric prior, letting color continuity supply depth hypotheses in textureless or far-field regions and reducing reliance on dense depth sensors.","The same Gaussian map could support downstream tasks the paper does not pursue, such as semantic segmentation or object-level editing, because each Gaussian is an individually optimizable primitive with spatial extent.","A natural next test is closed-loop and high-speed operation: the paper attributes degraded loop-closure and high-speed performance to the absence of loop closure and spherical harmonics, so adding those modules is a direct, falsifiable path to improvement.","Cross-modal radar localization on a Gaussian map suggests a testable extension: using the map as a shared representation for other sparse sensors in all-weather conditions."],"forward_implications":["A single LiDAR-camera pass can produce a photorealistic, differentiable map while simultaneously estimating the platform pose, removing the need for a separate offline reconstruction stage.","Because the map is continuous Gaussians rather than fixed-resolution point clouds, tracking and rendering are not bounded by LiDAR resolution, and areas between and beyond LiDAR beams can still be represented.","The rendering quality of the odometry output is close to the ground-truth-pose rendering, so the estimated trajectory is precise enough for appearance-based map use.","The Gaussian map carries enough geometric structure that a different modality such as mmWave radar can relocalize against it, suggesting the map is not only visually but geometrically consistent.","At 7.98 FPS with a 10 Hz LiDAR input, the pipeline approaches the LiDAR input rate, indicating potential for online use."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting representation and tile-based rendering approach that the whole system builds on.","marker":"[2]"},{"why":"Provides the covariance-based point-to-Gaussian alignment idea adapted for the front-end tracking module.","marker":"[4]"},{"why":"Provides the skybox initialization strategy used to handle unbounded scenes.","marker":"[9]"},{"why":"Serves as a Gaussian Splatting SLAM baseline and the source of the simplified color model without spherical harmonics.","marker":"[22]"},{"why":"Provides the open dataset and ground-truth trajectories used for evaluation and the closed-loop sequence.","marker":"[26]"},{"why":"Provides the graph-based LiDAR SLAM baseline and the localization routine used for cross-modal radar relocalization.","marker":"[27]"},{"why":"Serves as the feature-based visual SLAM baseline for trajectory comparison.","marker":"[28]"},{"why":"Serves as the neural-implicit LiDAR odometry baseline for trajectory comparison.","marker":"[3]"}],"fun_headline_variants":["3D Gaussians unify LiDAR-visual SLAM at 7.98 FPS","LiV-GS: first to fit sparse LiDAR into continuous Gaussian maps","Gaussian map SLAM goes beyond LiDAR field of view in real time","Cross-modal radar localization on a Gaussian SLAM map"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system relies on the assumption that a color-supervised Gaussian outside the LiDAR field of view is made reliable by pulling it toward the nearest LiDAR-measured Gaussian and treating split Gaussians as reliable after one round of back-end optimization; if that propagation does not preserve true surface geometry, the claimed reconstruction beyond the LiDAR field of view fails.","fun_headline_variants_meta":{"raw":{"variants":["3D Gaussians unify LiDAR-visual SLAM at 7.98 FPS","LiV-GS: first to fit sparse LiDAR into continuous Gaussian maps","Gaussian map SLAM goes beyond LiDAR field of view in real time","Cross-modal radar localization on a Gaussian SLAM map"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1398,"prompt_tokens":968,"completion_tokens":430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":348}},"tokens_in":584,"tokens_out":430,"duration_ms":4393,"temperature":1.0,"reasoning_tokens":348,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:50:13.589373+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the reconstructed surface in a region that LiDAR never observed and compare it against ground-truth geometry from a later dense LiDAR pass or a survey-grade scanner; if the conditional-Gaussian-split region deviates from the true surface by more than the in-field-of-view error, the central claim fails. A cheaper check is an ablation that removes the conditional Gaussian constraint and measures rendering quality on out-of-field-of-view image patches, since removing it should clearly degrade those regions if the constraint is doing the claimed work.","supporting_citations":[{"cited_title":"3D Gaussian Splatting for Real-Time Radiance Field Rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian Splatting representation and tile-based rendering approach that the whole system builds on."},{"cited_title":"Gaussian Splatting SLAM,","cited_arxiv_id":null,"evidence_quote":"Serves as a Gaussian Splatting SLAM baseline and the source of the simplified color model without spherical harmonics."},{"cited_title":"R3LIVE: A Robust, Real-Time, RGB-colored, LiDAR-Inertial-Visual tightly-coupled state Estimation and mapping package,","cited_arxiv_id":null,"evidence_quote":"Provides the open dataset and ground-truth trajectories used for evaluation and the closed-loop sequence."},{"cited_title":"A Portable Three-Dimensional LiDAR-based system for long-term and wide-area people behavior measurement,","cited_arxiv_id":null,"evidence_quote":"Provides the graph-based LiDAR SLAM baseline and the localization routine used for cross-modal radar relocalization."},{"cited_title":"ORB-SLAM3: An accurate open-source library for visual, visual–inertial, and multimap SLAM,","cited_arxiv_id":null,"evidence_quote":"Serves as the feature-based visual SLAM baseline for trajectory comparison."},{"cited_title":"NeRF-LOAM: Neural Implicit Representation for Large-scale incremental LiDAR odometry and mapping,","cited_arxiv_id":null,"evidence_quote":"Serves as the neural-implicit LiDAR odometry baseline for trajectory comparison."}],"review_version":1}