{"id":"b9d5b2bc-7102-4a6e-9484-e859f85b4947","arxiv_id":"2501.13417","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"GeomGS learns a per-Gaussian Geometric Confidence Score to weight LiDAR distance constraints, producing 3DGS maps that support weighted-ICP and image-refinement localization.","lead":"This paper combines LiDAR point clouds with 3D Gaussian Splatting using a per-point confidence score, then localizes a robot by weighted point-cloud matching plus image-based refinement. The authors report better rendering, geometry, and localization on KITTI and KITTI-360, but the localization evaluation appears to test on frames already used to build the map.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Localization evaluation is circular: query scans are drawn from the 100-frame mapping set, so WICP aligns each query to its own LiDAR points in the map.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the localization evaluation does not exclude query LiDAR scans from the accumulated map prior P. This is the single most damaging issue because it directly invalidates the paper's central claim of state-of-the-art localization performance. The paper's reported geometric accuracy (Table II) is also measured against the accumulated LiDAR points P, which are the fitting target of the constraint losses, so high F-scores partly reflect the loss design rather than independent geometric quality. However, the localization flaw is more consequential: if the query scan is in the map, the weighted ICP step has access to the query's own points, making the registration problem near-trivial regardless of whether GCS provides genuine geometric weighting. The image refinement component is similarly compromised because query images are drawn from the training set used to optimize L_rgb. A held-out evaluation is a minimal, well-defined fix that would settle the matter, and until it is performed the REJECT verdict stands. I considered whether the probabilistic distance constraint (Eq. 9) is itself a second fatal flaw, but it is secondary: even if the loss is simply a reweighted distance, the method's contribution could still be judged on the mapping quality if the evaluation were fair. For the localization claim, no amount of internal consistency rescues a circular test. The concrete check I propose directly addresses the circularity: rebuild the map without the query frames and re-measure errors. If the method passes that test, the paper's main claim has a sound empirical basis; if it fails, the claim is unsupported. I therefore recommend no change to the reader's REJECT verdict, with the understanding that a corrected evaluation could lead to a different outcome.","tokens_in":10989,"tokens_out":2890,"duration_ms":28153,"concrete_test":"Hold out the query frames from mapping: build the GeomGS map using only frames 1-90 of the same 100-frame KITTI/KITTI-360 segment, and localize with frames 91-100 (or, equivalently, exclude every 10th frame from the map while keeping the other 90 frames for mapping). Re-run the Table III comparison with identical initial errors, iteration counts, and baselines, reporting rotation and translation errors for ICP, WICP, iNeRF, and Ours. If the mean translation error of WICP or Ours degrades substantially relative to Table III (e.g., from about 0.54 m to over 2 m), the localization benefit of GeomGS is not demonstrated; if errors remain comparable, the circularity concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central localization claim in Table III is supported by an evaluation protocol that appears to test the method on the very data used to build the map. Section IV-A states that mapping uses 100 consecutive images from KITTI/KITTI-360, and Eq. 5 builds the accumulated LiDAR prior P as the union of all transformed scans, P = ∪_i (T_i P_i). Section IV-D then selects test cases from every 10th sequence out of those 100 sequences, without stating that those sequences are excluded from mapping. Under the natural reading, each query scan Q_p is a subset of P, so the weighted ICP step (Eqs. 14-16) minimizes the distance from Q_p to a map that already contains Q_p's own points. The reported translation/rotation errors therefore measure registration against a map containing the query, not localization in an unseen environment. The same circularity affects the image-refinement stage: query images are likely training views used in L_rgb (Eq. 10), so rendering at a near-correct pose is artificially easy to match. If the authors intended a held-out split, it is not described; as written, the experimental design does not support the abstract's claim of state-of-the-art localization performance. A proper held-out evaluation is required before the central claim can be assessed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeomGS, a 3D Gaussian Splatting variant that initializes Gaussians from accumulated LiDAR point clouds and introduces a per-Gaussian Geometric Confidence Score (GCS). GCS is fitted to an asymmetric sigmoid of the distance to the nearest LiDAR point and is used to reweight a probabilistic distance loss. A localization module alternates GCS-weighted ICP against the Gaussian map with iNeRF-style image refinement. Experiments on KITTI and KITTI-360 report image quality, geometric fidelity, and localization accuracy.","tokens_in":11338,"tokens_out":6461,"duration_ms":56225,"significance":"Accurate metric-scale maps for localization are a relevant problem, and combining LiDAR priors with 3DGS is timely. The paper presents a concrete system with comparisons to several 3DGS variants and a localization pipeline that integrates LiDAR and photometric cues. However, the evaluation protocol as described does not establish the claimed state-of-the-art localization or geometric accuracy, because the query LiDAR scans and test images appear to be part of the mapping set. If the method were re-evaluated on a genuinely held-out split, the underlying idea of GCS-weighted ICP with image refinement could still be of interest, but the central claims currently rest on an evaluation that measures self-registration rather than generalization.","major_comments":[{"comment":"The localization evaluation is circular. Mapping uses 100 consecutive images and builds the accumulated LiDAR prior P = union_i (T_i P_i) from the poses of those images (Eq. 5). Section IV-D then selects test cases from every 10th sequence out of those 100 sequences without stating that those sequences are excluded from mapping. Under the natural reading, each query scan Q_p is a subset of P, so the weighted ICP (Eqs. 14-16) aligns the query to a map that already contains the query's own points. The errors in Table III therefore measure self-registration, not localization against an unseen map. The image-refinement stage has the same problem if the query images are among the images used to train the photometric loss. Please re-run the evaluation with a held-out split, such as a separate sequence or temporally disjoint scans, and report exactly which frames and scans are used for mapping versus testing.","section":"Section IV-A/IV-D, Table III"},{"comment":"The geometric quality metrics in Table II are computed between the optimized Gaussians G and the accumulated LiDAR points P, which is the very target used in the distance losses L_geom and L_prob (Eqs. 7-9). A method that is explicitly constrained to move Gaussians toward P will trivially score well on Chamfer distance and F-score against P. This does not demonstrate that the map is geometrically accurate in an independent sense; it only shows that the constraint was enforced. Please evaluate against held-out LiDAR scans, a ground-truth mesh, or at least a separate validation LiDAR scan that was not used in mapping.","section":"Section IV-C, Table II"},{"comment":"The GCS is not learned from an independent geometric signal: Eq. 8 fits gamma_i to a hand-set asymmetric sigmoid of d_i, and Eq. 9 then uses gamma_i to reweight the same d_i in the distance loss. As written, the 'probabilistic distance constraint' is a deterministic reweighting of the nearest-neighbor distance by a fixed function of that distance. The paper should clarify what is actually learned beyond the hand-set sigmoid parameters k and d, and justify why this constitutes a probabilistic model rather than a heuristic robust weighting. If the intended contribution is only a heuristic weighting, the terms 'probabilistic' and 'Geometric Confidence Score' as a learned parameter should be revised accordingly.","section":"Section III-B, Eqs. (6)-(9)"},{"comment":"The image-quality evaluation is described as 'testing on scenes sampled every 8 frames' from the same 100-frame sequence used for mapping, with no statement that these frames are held out during training. If the training set includes all 100 frames, the PSNR/SSIM/LPIPS results in Table I are self-reconstruction scores rather than novel-view synthesis results. Please specify the exact training/test image split and, if the current protocol does not hold out test frames, retrain on the remaining frames and report the corresponding numbers.","section":"Section IV-B, Table I"}],"minor_comments":[{"comment":"There are several typos and formatting errors, including 'QUANTITAIVE' in Table I, 'PERFORMACE' in Table II, and the footnote '∗is that the corresponding author' in Section I.","section":"Throughout"},{"comment":"In Eq. (10), the L1 term uses I_a (appearance-model image) while the SSIM term uses I_r (rendered image); it would be clearer to state whether the L1 loss is intentionally computed against the appearance-model output rather than the standard rendered image.","section":"Eq. (10)"},{"comment":"In Eq. (14), W = diag(gamma) uses GCS values, but it is not specified whether these weights are applied to the source (query LiDAR) points, the target (Gaussian map) points, or both; since the query scan has no GCS, the intended construction of S, T, and W should be stated explicitly.","section":"Eq. (14)"},{"comment":"The paper uses both '100 consecutive images' and 'every 10th sequence out of 100 sequences'; it should clarify whether a 'sequence' is a single frame or a sub-sequence, since this affects the number of localization test cases.","section":"Section IV-A"},{"comment":"No standard deviations or number of runs are reported for the localization experiments, so it is unclear whether the differences between methods are statistically significant.","section":"Table III"},{"comment":"Some references are incomplete or inconsistently formatted, such as [33] (missing venue) and [34] (arXiv preprint without a clear publication status).","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The main blocker is the evaluation protocol. The paper appears to test localization and image quality on the same data used to build the map, which invalidates the central claims as written. The geometric metrics are also computed against the fitting target. These issues are fixable with a genuine held-out split, so I would not reject outright, but I would require the authors to report explicit mapping/test splits for both LiDAR and images, and to verify that the reported gains persist under that protocol. If the authors cannot produce such experiments, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core mapping idea has some merit: adding a per-Gaussian confidence score (GCS) to down-weight LiDAR constraints on unreliable points (sky, tall buildings) is a sensible fix for DrivingGaussian's crude distance loss, and the weighted-ICP-with-image-refinement pipeline is a natural way to use the resulting map. The GCS parameterization is not verbatim in prior work, and the paper is honest about its motivation.\n\nThat said, the central localization claim is unsupported. Section IV-A says mapping uses 100 consecutive images from KITTI/KITTI-360, and Eq. 5 builds the LiDAR prior P as the union of all transformed scans. Section IV-D then picks test cases from every 10th of those 100 sequences, with no statement that they are excluded from mapping. Under the natural reading, each query scan Qp is a subset of P, so the weighted ICP in Eqs. 14-16 aligns the query to the map that already contains the query's own points. The same problem hits the image-refinement stage: the query images are almost certainly training views, so rendering at a near-correct pose is artificially easy. Table III therefore does not measure localization at all; it measures self-registration. The stress-test note is right on the money.\n\nThere is a second, related circularity in the geometry evaluation. Table II computes F-score and Chamfer distance between the fitted Gaussians G and the accumulated LiDAR points P—the very target used in the loss. That tells you how well the model fits its own training data, not how geometrically accurate the reconstruction is against held-out ground truth. The paper can legitimately report this as a fitting diagnostic, but not as a reconstruction accuracy claim.\n\nThe 'probabilistic' framing in Eq. 9 is also thinner than advertised: GCS is trained to match a hand-set sigmoid of distance (Eq. 8) and then used to reweight the same distances. That is a reweighted distance loss, not a principled probabilistic model. Minor in isolation, but it adds to the overclaiming.\n\nSo: the mapping part might be salvageable, but the evaluation needs a genuine held-out split—both for LiDAR scans and images—before any localization number can be trusted. As written, the paper overreaches. A serious referee should be willing to look past the flaws and demand that fix, because the GCS idea could be useful. I would send it to review with a strong request for held-out experiments, but I would not cite it in its current form.","headline":"The localization evaluation is circular—query scans come from the same 100 frames used to build the map—so the headline accuracy numbers measure self-registration, not localization.","tokens_in":738,"tokens_out":749,"would_cite":false,"duration_ms":32476,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a Geometric Confidence Score to LiDAR-seeded Gaussian splatting yields metric-scale maps accurate enough for scan-plus-image robot localization.","keywords":["3D Gaussian Splatting","LiDAR-guided mapping","Geometric Confidence Score","robot localization","weighted ICP","pose estimation","metric-scale reconstruction","autonomous driving"],"falsifier":"Rebuild each map without the LiDAR scans belonging to the test frames, rerun the paper's localization protocol, and compare rotation and translation errors; if the errors grow substantially, the reported localization accuracy depends on the query scan being contained in the map.","tokens_in":10804,"feed_emoji":"📍","tokens_out":8968,"duration_ms":73687,"temperature":0.7,"pith_summary":"GeomGS addresses the problem that 3D Gaussian Splatting maps often look right but are metrically wrong, because their geometry comes from structure-from-motion rather than from physical range measurements. The paper proposes to seed the Gaussians with accumulated LiDAR points and to give each Gaussian a Geometric Confidence Score that says how structurally reliable it is, optimized together with the Gaussians under a probabilistic distance constraint so the map is pulled toward LiDAR geometry where it matters while still rendering well. On top of this map, the paper builds a localization method that alternates confidence-weighted LiDAR scan-to-map alignment with image-based pose refinement. The intended payoff is a renderable, metric-scale map and a pose estimator a robot can use in autonomous driving. The paper supports this claim with image-quality, geometric-accuracy, and pose-error experiments on urban-driving benchmarks.","feed_headline":"Per-point confidence scores turn LiDAR splats into robot maps","feed_subtitle":"A geometric confidence score guides weighted scan alignment plus image-based pose refinement.","key_machinery":"The load-bearing object is the Geometric Confidence Score $\\gamma \\in (0,1)$, computed toward an asymmetric sigmoid $\\sigma_{\\mathrm{asym}}(x) = 1/(1+e^{k(x-d)})$ with $k=20$ and $d=0.9$, where $x$ is the distance $d_i = \\min_{p\\in P}\\|g_i-p\\|_2$ from Gaussian $g_i$ to the accumulated LiDAR prior $P$. Two losses carry the geometry: $\\mathcal{L}_{\\mathrm{geom}}$ drives $\\gamma_i$ toward $\\sigma_{\\mathrm{asym}}(d_i)$, and $\\mathcal{L}_{\\mathrm{prob}}$ applies the cost $(1/N)\\sum_i (\\ln(1-\\gamma_i) + d_i/(1-\\gamma_i))$, which downweights Gaussians that exist mainly for rendering. The same GCS is reused at localization time as the diagonal weight matrix in weighted ICP, linking mapping quality directly to pose estimation.","core_discovery":"The paper's central claim is that LiDAR can be fully integrated into 3D Gaussian primitives rather than merely used as initialization or a soft pull. GeomGS introduces the Geometric Confidence Score, a per-Gaussian parameter optimized to be close to an asymmetric sigmoid of the distance to the nearest accumulated LiDAR point, and a probabilistic distance loss that assigns higher weight to high-confidence Gaussians. The result is claimed to be a map that preserves metric scale and structure without sacrificing photometric quality. On that map the paper proposes a localization pipeline in which a query LiDAR scan is aligned by weighted ICP using GCS as per-point weights, the pose is then refined by minimizing the photometric error between a rendered image and the query image, and the two steps are iterated. Across its urban-driving benchmarks, the paper reports the lowest Chamfer distance and highest F-score among the compared methods, improved PSNR/SSIM/LPIPS relative to the baselines it compares, and lower rotation and translation errors than plain ICP, weighted ICP alone, or image-only refinement.","pith_inferences":["The paper's localization protocol selects every 10th of 100 frames as test cases but never states that these frames are excluded from the accumulated LiDAR prior built over all poses; if they are included, the reported errors partly measure registration to the query's own points rather than localization within an unseen map.","A natural test the paper does not run is to rebuild the map from the remaining 90 scans and repeat the localization table; that would quantify how much of the claimed accuracy survives genuine map novelty.","The GCS could be exported beyond localization, for example as a reliability mask for pruning unstable Gaussians or for detecting changed regions in a revisited scene.","The probabilistic distance loss's behavior under denser or noisier LiDAR is untested; perturbing the input scans would show whether the structural gains persist across sensor characteristics."],"forward_implications":["Maps produced by GeomGS carry metric scale and structural detail, so downstream robotics tasks that need true geometry, not just pretty rendering, can use them directly.","The same per-point confidence map that guides mapping also serves as ICP weighting, so geometric reliability discovered during mapping transfers to localization without extra learned components.","Alternating weighted scan alignment with image-based refinement covers each method's failure mode: image refinement corrects scan misalignment, while scan alignment fixes large pose errors that image-only refinement misses.","Because the geometry losses only pull high-confidence Gaussians toward LiDAR, regions such as sky and tall buildings can remain rendering-oriented without distorting the surveyed structure."],"supporting_citations":[{"why":"Defines 3D Gaussian Splatting, the base representation GeomGS extends with LiDAR-guided geometry.","marker":"[14]"},{"why":"Introduces the simple distance constraint between Gaussian points and LiDAR points that GeomGS targets and attempts to improve.","marker":"[7]"},{"why":"Supplies the probabilistic distance formulation behind the GCS-weighted constraint.","marker":"[8]"},{"why":"Defines ICP, the alignment algorithm GeomGS generalizes to weighted ICP via GCS weights.","marker":"[9]"},{"why":"Provides the image-based pose refinement that GeomGS alternates with weighted ICP, and serves as a baseline.","marker":"[27]"},{"why":"Gives the F-score and Chamfer-distance evaluation protocol used to measure geometric accuracy against accumulated LiDAR points.","marker":"[24]"},{"why":"Supplies the structure-from-motion initialization baseline that GeomGS replaces with LiDAR points.","marker":"[1]"},{"why":"Supplies the appearance model used in the photometric loss to handle brightness variation.","marker":"[21]"},{"why":"Supplies the scale loss that prevents overlapping Gaussians.","marker":"[36]"},{"why":"Supplies the perceptual loss used to preserve feature-level detail.","marker":"[37]"}],"fun_headline_variants":["Per-point confidence scores turn LiDAR splats into robot maps","Probabilistic LiDAR integration gives Gaussians structural confidence","Geometric confidence scores make splats localize robots","LiDAR-weighted Gaussians: sharper maps, better poses","GCS: tying LiDAR to Gaussian mapping for precise localization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation's load-bearing premise is that the every-10th query scan is not part of the accumulated LiDAR prior P; the paper never states this exclusion, so the reported localization errors may measure alignment to a map containing the query's own points.","fun_headline_variants_meta":{"raw":{"variants":["Per-point confidence scores turn LiDAR splats into robot maps","Probabilistic LiDAR integration gives Gaussians structural confidence","Geometric confidence scores make splats localize robots","LiDAR-weighted Gaussians: sharper maps, better poses","GCS: tying LiDAR to Gaussian mapping for precise localization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1443,"prompt_tokens":962,"completion_tokens":481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":399}},"tokens_in":578,"tokens_out":481,"duration_ms":6112,"temperature":1.0,"reasoning_tokens":399,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:58:06.505281+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rebuild each map without the LiDAR scans belonging to the test frames, rerun the paper's localization protocol, and compare rotation and translation errors; if the errors grow substantially, the reported localization accuracy depends on the query scan being contained in the map.","supporting_citations":[{"cited_title":"Drivinggaussian: Composite gaussian splatting for surrounding dy- namic autonomous driving scenes,","cited_arxiv_id":null,"evidence_quote":"Introduces the simple distance constraint between Gaussian points and LiDAR points that GeomGS targets and attempts to improve."},{"cited_title":"Robust 6d object pose estimation by learning rgb-d features,","cited_arxiv_id":null,"evidence_quote":"Supplies the probabilistic distance formulation behind the GCS-weighted constraint."},{"cited_title":"A method for registration of 3-d shapes,","cited_arxiv_id":null,"evidence_quote":"Defines ICP, the alignment algorithm GeomGS generalizes to weighted ICP via GCS weights."},{"cited_title":"Points2nerf: Generating neural radiance fields from 3d point cloud,","cited_arxiv_id":null,"evidence_quote":"Gives the F-score and Chamfer-distance evaluation protocol used to measure geometric accuracy against accumulated LiDAR points."},{"cited_title":"Vastgaussian: Vast 3d gaussians for large scene reconstruction,","cited_arxiv_id":null,"evidence_quote":"Supplies the appearance model used in the photometric loss to handle brightness variation."},{"cited_title":"Mixture of volumetric primitives for efficient neural rendering,","cited_arxiv_id":null,"evidence_quote":"Supplies the scale loss that prevents overlapping Gaussians."}],"review_version":1}