{"id":"1d78f578-68cf-4afa-81e5-240151eda4ee","arxiv_id":"2505.10601","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SRMamba applies a Mamba state-space model to range-view LiDAR super-resolution, adding Hough-voting projection and hole compensation, and reports improved Chamfer Distance and IoU over Transformer and attention baselines on KITTI-360 and nuScenes.","lead":"SRMamba is a neural network that turns sparse LiDAR point clouds into denser ones by converting them to depth images, filling blank stripes, upsampling the images with a Mamba-based model, and projecting back to 3D. It reports better reconstruction metrics than three recent methods on two driving datasets, but the evaluation uses a synthetic downsampling setup and no code is released.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative comparison rests on synthetic 4x beam downsampling; transfer to genuine low-beam LiDAR is untested, so the central superiority claim remains conditional.","rationale":"The reader's weakest assumption is the same one I would flag. Section 4.1 explicitly uses 4x downsampling to simulate sparse input, and all tables and figures compare on that synthetic task, while the paper's central claim is about real LiDAR super-resolution. A real low-beam sensor is not simply a subset of a high-beam sensor's scan lines: elevation angles, beam divergence, range noise, and occlusion patterns differ, so the synthetic benchmark cannot by itself establish the deployment claim. The architecture and internal comparison are reasonable, and I am not objecting to them; I am only requiring a transfer check before accepting the abstract's wording. This supports the existing CONDITIONAL verdict without changing it.","tokens_in":12764,"tokens_out":6390,"duration_ms":69149,"concrete_test":"Re-run the KITTI-360 comparison using a VLP-16-accurate simulation instead of uniform 4x downsampling: keep only beams whose elevation angles match a real VLP-16, apply that sensor's range noise model, and retrain/evaluate SRMamba and TULIP on this input; if the reported IoU/CD gains over TULIP shrink materially or invert, the transfer claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition is stated in Section 4.1: 'the two datasets are processed with 4 times downsampling to simulate sparse point cloud inputs.' Every quantitative comparison in Tables 1-2 and Fig. 10 is on this synthetic task: KITTI-360's HDL-64E is reduced to 16 beams and nuScenes's HDL-32E to 8 beams by row dropping. For the abstract's central claim that SRMamba is significantly superior for LiDAR point cloud super-resolution and restores 3D structure under novel views, this simulation must faithfully represent a genuine low-beam LiDAR input. It does not: real 16-beam sensors such as the VLP-16 have their own elevation angles, nonuniform beam spacing, range noise, and different occlusion patterns, whereas dropping every fourth beam of a 64-beam sensor yields a subset of the dense sensor's angles with noise-free geometry. Additionally, the 'novel view' advantage is only shown qualitatively in Fig. 9; no quantitative metric is reported for off-projection viewpoints. The reported IoU, CD, and MAE margins therefore establish superiority on synthetic downsampling, not yet on the real sparse-sensor setting the paper motivates.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SRMamba, a range-image-based LiDAR point cloud super-resolution method. The pipeline projects a sparse point cloud to a range image, applies Hough Voting and a Hole Compensation mechanism to reduce horizontal holes, then uses a Mamba-based asymmetric U-Net with multi-directional scanning to produce a high-resolution range image before back-projecting to 3D. Experiments on KITTI-360 and nuScenes with simulated 4x beam downsampling compare SRMamba against Cas-ViT, SwinIR, and TULIP using IoU, Chamfer Distance, and MAE. The authors report consistent quantitative gains and qualitative improvements, and claim that the method better preserves 3D structure under novel viewpoints.","tokens_in":12980,"tokens_out":2542,"duration_ms":25553,"significance":"If the claims hold, the paper would make a useful contribution: it combines a state-space model with range-view super-resolution for LiDAR point clouds, addresses hole artifacts with a simple preprocessing mechanism, and reports experiments on two public datasets with a published baseline comparison. The explicit failure-case discussion is also commendable. However, the significance is currently tempered by the evaluation protocol: the central superiority claim rests on synthetic beam downsampling, the novel-view advantage has no quantitative evidence, and the reported margins lack error bars. These issues make the contribution plausible but not yet established.","major_comments":[{"comment":"The load-bearing evaluation condition is the statement that 'the two datasets are processed with 4 times downsampling to simulate sparse point cloud inputs.' All quantitative results (Tables 1-2 and Fig. 10) are obtained on this synthetic task: KITTI-360's HDL-64E is reduced to 16 beams and nuScenes's HDL-32E to 8 beams by row dropping. Real low-beam LiDAR sensors have different elevation angles, nonuniform beam spacing, and different noise and occlusion characteristics, so a subset of a dense sensor's beams is not the same as a genuine sparse sensor input. Without an experiment on real sparse LiDAR data, or at minimum a much stronger argument that the synthetic task transfers, the abstract's claim of 'significant superiority' remains conditional on the simulation being faithful.","section":"Section 4.1 (Dataset) and Tables 1-2"},{"comment":"The paper's central motivation and abstract emphasize reconstruction under novel views, but the novel-view advantage is demonstrated only qualitatively in Fig. 9. No quantitative metric is reported for viewpoints off the projection plane, even though the introduction states that the method 'address[es] the key challenge of recovering the 3D spatial structure of point clouds from novel views.' A quantitative evaluation at rotated viewpoints (e.g., CD or F-score after transforming the reconstruction) is needed to support this claim.","section":"Section 4.4 (Qualitative Evaluation, Fig. 9) and abstract"},{"comment":"The quantitative comparison reports only single validation-split numbers without error bars, repeated runs, or any statistical significance test. Because the reported margins are moderately sized (e.g., IoU 0.4548 vs 0.4152 for TULIP on KITTI-360), a reader cannot tell whether these differences are stable or within run-to-run noise. The validation-only protocol (Section 4.1) also leaves open the risk of overfitting to the selected scans; held-out test splits or cross-validation would strengthen the claim.","section":"Section 4.4.2 and Tables 1-2"}],"minor_comments":[{"comment":"The abstract says experiments were conducted on 'SemanticKITTI,' but Section 4.1 uses KITTI-360. This is a factual mismatch that should be corrected.","section":"Abstract and Section 4.1"},{"comment":"The text states that SRMamba 'achieves better performance in terms of all metrics,' but on KITTI-360 SRMamba-L has MAE 0.0048 while SRMamba-T has MAE 0.0044, so the larger model is worse on MAE. The sentence should be qualified.","section":"Section 4.4.2 and Table 1"},{"comment":"The phrase '600 ephemeral sessions' is unclear; it presumably means 600 training epochs. Also, the training/test protocol does not state whether the baselines (Cas-ViT, SwinIR, TULIP) were trained under identical conditions, which is important for a fair comparison.","section":"Section 4.3 (Experimental Details)"},{"comment":"The notation for Δb and 𝜑b says they are 1D arrays of length H, but in Eq. (6) they are used as though each point has a corresponding value. Please clarify the indexing and whether the argmin is over the beam list.","section":"Section 3.2.2, Eq. (6)"},{"comment":"The depth ablation shows SRMamba-M with MAE 0.0055, worse than SRMamba-S (0.0044) and SRMamba-T (0.0044), while CD and IoU improve monotonically. The text says the results 'validate the effectiveness of the proposed method for multi-layer feature extraction,' but the trend is not monotonic in all metrics and should be discussed.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The reviewer agrees with the stress-test assessment: the synthetic beam-downsampling assumption is the main correctness-risk point. The paper would be substantially strengthened by experiments on genuine sparse LiDAR data (e.g., a small real 16-beam dataset or a carefully justified sensor-transfer experiment) and by quantitative novel-view metrics. The lack of error bars is concerning for a paper whose headline contribution is empirical superiority. The fit with the journal is appropriate; the issues are fixable within the manuscript's scope, so major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the one-liner: this is a solid engineering paper, not a breakthrough. The genuinely new piece is the Hough-voting projection plus hole compensation preprocessing for range-view LiDAR super-resolution, combined with a VMamba-style four-direction scanning U-Net. That combination doesn't appear in the prior work I know, and the method is coherent. The authors also include a failure-case section, which is honest.\n\nWhat's good: the experiments are on two public datasets with standard metrics, and the qualitative figures, especially the novel-view ones, show a real improvement in structure preservation over TULIP, SwinIR, and Cas-ViT. The ablation study on preprocessing and network depth is informative. There's no sign of target fitting; the preprocessing uses sensor calibration, so circularity is not a concern.\n\nThe soft spots are real but proportionate. The biggest one is the evaluation protocol: 4x row-dropping of a dense scan is not the same as a genuine 16- or 8-beam sensor, which has different elevation angles, nonuniform spacing, and noise. The paper motivates the task with real low-beam LiDAR but never tests on it. The novel-view advantage is only qualitative. Also, the abstract says SemanticKITTI while the experiments use KITTI-360—an error that must be fixed. Table 3 shows the network's marginal gain over the preprocessing alone is small: Hough Voting alone lifts TULIP from 0.4152 to 0.4255 IoU, and SRMamba-T with Hough Voting reaches 0.4369, just 0.002 below the full model. No error bars, no repeated runs, no code release. RangeLDM is cited but not compared, which is a missed opportunity.\n\nThe paper is for people working on low-cost LiDAR densification for autonomy. It deserves a serious referee rather than a desk reject. I'd want the authors to fix the dataset-name error, report variance, address the simulation gap explicitly, and ideally add a quantitative metric for novel views. The core idea is worth engaging with.","headline":"Solid but incremental range-view LiDAR SR; the Hough-voting preprocessing is the real novelty, but the evaluation's synthetic downsampling and missing error bars keep the superiority claim conditional.","tokens_in":13532,"tokens_out":2903,"would_cite":true,"duration_ms":27452,"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":"SRMamba reconstructs 4x denser LiDAR point clouds from sparse scans using a state-space range-image network, and reports better geometry than attention-based baselines.","keywords":["LiDAR super-resolution","point cloud upsampling","range image","state space model","Mamba","Hough voting","hole compensation","novel view"],"falsifier":"Record a scene simultaneously with a high-beam LiDAR and a real low-beam LiDAR, train SRMamba on the genuine low-beam scans, and compare its upsampled cloud to the high-beam ground truth; if the IoU and Chamfer Distance margins over the Transformer baselines shrink to noise, or the oblique-view geometry drifts, the synthetic-downsampling premise fails.","tokens_in":12538,"feed_emoji":"🚗","tokens_out":8140,"duration_ms":75752,"temperature":0.7,"pith_summary":"This paper argues that a 4x denser LiDAR point cloud can be reconstructed from a sparse one by turning the problem into range-image super-resolution and solving it with a state-space (Mamba) network instead of a Transformer. The proposed architecture, SRMamba, projects the sparse cloud into a range image, repairs horizontal and discrete holes with Hough voting and a hole-compensation window, then upsamples the image with an asymmetric U-Net built from visual state-space blocks that scan the feature map in four directions. On two public automotive LiDAR datasets, the paper reports that SRMamba beats three attention-based baselines on IoU, Chamfer Distance, and MAE, with the clearest qualitative gains at oblique viewpoints where the baselines produce drift, noise, and collapsed structures. If the claim holds, dense high-beam-like geometry can be approximated from low-beam scans at linear rather than quadratic sequence-modeling cost.","feed_headline":"SRMamba turns sparse LiDAR scans into 4x denser point clouds","feed_subtitle":"A Mamba-based U-Net with hole-filling beats attention baselines, especially at oblique views.","key_machinery":"The central machinery is a range-image preprocessing pipeline followed by a visual state-space (VSS) block inside an asymmetric U-Net. A spherical projection maps each LiDAR point to a range image using per-beam vertical offsets; Hough voting estimates projected coordinate offsets and a hole-compensation window fills the remaining empty pixels before the network sees the image. The network is an asymmetric U-Net whose stages are VSS blocks: the SS2D module scans the feature map left-to-right, right-to-left, top-to-bottom, and bottom-to-top, giving a global receptive field with linear-time state updates instead of quadratic attention. Skip connections provide multi-scale feature fusion, PixelShuffle upsamples the range image to four times its height, and the resulting depth map is back-projected to 3D space.","core_discovery":"SRMamba's central claim is that range-view LiDAR super-resolution fails on novel viewpoints because projection leaves horizontal holes and because windowed attention cannot capture long-range 3D spatial dependencies. The paper proposes to fix both: Hough voting plus hole compensation repairs the range image before learning, and a visual state-space backbone with multi-directional scanning replaces attention. On the 64-beam urban benchmark, SRMamba-L reaches IoU 0.4548 and Chamfer Distance 0.0940, improving on the Transformer baseline by 9.5% and 24.3% respectively; on the 32-beam driving benchmark it reaches IoU 0.3482 and Chamfer Distance 0.9620, with clear margins over all three compared baselines. The paper further claims that the reconstructed clouds stay spatially consistent in oblique views, recovering structures such as vehicle roofs, wall outlines, and truck contours that baselines miss or clutter with noise.","pith_inferences":["The paper does not test temporal fusion; if the synthetic-downsampling assumption holds, the same range-image pipeline could plausibly average consecutive low-beam scans to fill holes before super-resolution, which could further reduce drift.","A matched-capacity ablation with the same U-Net depth and training schedule but a Transformer block instead of the SS2D block, both fed the same Hough-voting inputs, would isolate how much of the gain comes from the Mamba backbone rather than the hole-filling preprocessing.","Because the output is a range image, the method should adapt to other spinning LiDAR beam counts and possibly to solid-state sensors whose scans can be re-projected, but real-sensor evaluation is needed to confirm transfer.","The paper reports reconstruction metrics only; testing SRMamba-upsampled clouds in downstream detection or segmentation would show whether the geometric gains translate into perception gains."],"forward_implications":["A 4x denser point cloud can be generated from a sparse LiDAR scan with a sequence model whose cost grows linearly with image length, making high-density emulation feasible on cheaper sensors.","Repairing projection holes before learning matters: the ablation shows that adding Hough voting alone lifts the Transformer baseline's IoU from 0.4152 to 0.4255 on the 64-beam benchmark and cuts Chamfer Distance from 0.1241 to 0.1068.","Multi-directional scanning helps recover 3D structure at novel oblique viewpoints, reducing the coordinate drift and pseudo-points that windowed attention produces.","The asymmetric U-Net adapts to different beam counts, super-resolving both 16x1024 and 8x1024 range images with the same architecture, which the paper presents as evidence of multi-beam generalization.","The method maintains higher accuracy at long range than the baselines, with the largest measured advantage in the 40-50 meter interval."],"supporting_citations":[{"why":"The Transformer-based LiDAR upsampler that serves as the primary baseline; its IoU and Chamfer Distance are the numbers SRMamba must beat.","marker":"[38]"},{"why":"The visual state-space model whose SS2D scanning module SRMamba adapts for long-range range-image dependencies.","marker":"[20]"},{"why":"The selective state-space sequence model that provides the linear-time scanning formalism behind the VSS block.","marker":"[41]"},{"why":"The 64-beam urban LiDAR dataset whose scans are 4x downsampled to simulate sparse inputs for the main experiments.","marker":"[48]"},{"why":"The 32-beam autonomous-driving LiDAR dataset used for the sparser 8-beam evaluation.","marker":"[49]"},{"why":"A convolutional additive self-attention vision transformer used as one of the comparative baselines.","marker":"[53]"},{"why":"An image-restoration Swin Transformer baseline that the paper compares against after range-image projection.","marker":"[54]"}],"fun_headline_variants":["Mamba-based LiDAR SR fills holes, beats attention at oblique views","SRMamba mends range-view holes to densify LiDAR point clouds","Hough voting + Mamba fixing holes for sharper LiDAR upsampling","New Mamba model recovers LiDAR structures missed by attention","SRMamba: Mamba backbone super-resolves sparse LiDAR, even from new angles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that taking a dense LiDAR scan and removing three of every four scan lines faithfully mimics what a genuinely sparse low-beam LiDAR would see, so the reported gains may not transfer to real 16-beam or 8-beam sensors.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-based LiDAR SR fills holes, beats attention at oblique views","SRMamba mends range-view holes to densify LiDAR point clouds","Hough voting + Mamba fixing holes for sharper LiDAR upsampling","New Mamba model recovers LiDAR structures missed by attention","SRMamba: Mamba backbone super-resolves sparse LiDAR, even from new angles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00095,"raw_usage":{"total_tokens":4065,"prompt_tokens":971,"completion_tokens":3094,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":2994}},"tokens_in":587,"tokens_out":3094,"duration_ms":21310,"temperature":1.0,"reasoning_tokens":2994,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:10:28.513676+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a scene simultaneously with a high-beam LiDAR and a real low-beam LiDAR, train SRMamba on the genuine low-beam scans, and compare its upsampled cloud to the high-beam ground truth; if the IoU and Chamfer Distance margins over the Transformer baselines shrink to noise, or the oblique-view geometry drifts, the synthetic-downsampling premise fails.","supporting_citations":[{"cited_title":"Tulip: Transformer for upsampling of lidar point clouds,","cited_arxiv_id":null,"evidence_quote":"The Transformer-based LiDAR upsampler that serves as the primary baseline; its IoU and Chamfer Distance are the numbers SRMamba must beat."},{"cited_title":"Vmamba: Visual state space model,","cited_arxiv_id":null,"evidence_quote":"The visual state-space model whose SS2D scanning module SRMamba adapts for long-range range-image dependencies."},{"cited_title":"Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d,","cited_arxiv_id":null,"evidence_quote":"The 64-beam urban LiDAR dataset whose scans are 4x downsampled to simulate sparse inputs for the main experiments."},{"cited_title":"Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking,","cited_arxiv_id":null,"evidence_quote":"The 32-beam autonomous-driving LiDAR dataset used for the sparser 8-beam evaluation."},{"cited_title":"Swinir: Image restoration using swin transformer,","cited_arxiv_id":null,"evidence_quote":"An image-restoration Swin Transformer baseline that the paper compares against after range-image projection."}],"review_version":1}