{"id":"9d73c141-6f58-4747-b17e-ef7efa7b5bb9","arxiv_id":"2508.12986","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Mamba-based point upsampling network is introduced for single-photon point clouds, showing improved reconstruction metrics on noise-augmented benchmarks and visually cleaner output on real LiDAR data.","lead":"The authors propose SPU-MAMBA, a deep network based on the Mamba state-space model, to upsample sparse, distorted point clouds from single-photon LiDAR. It reports improved Chamfer distance and Hausdorff distance over existing point upsampling methods on standard benchmarks with simulated depth noise.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'first to establish the upsampling framework for single-photon sensing' claim is not supported: no single-photon-specific reconstruction or super-resolution method is included as a baseline, so the novelty and practical-utility claims rest on an incomplete comparison.","rationale":"The reader's weakest assumption focuses on the validity of the synthetic Gaussian depth-offset noise model. I agree that this is a real concern, but I identify a more load-bearing issue: the novelty claim of being 'first to establish the upsampling framework for single-photon sensing' is not supported by the experimental protocol, because no single-photon-specific reconstruction method is included as a baseline. The paper itself (Sec. II.B) lists such methods, so the omission is not an artifact of an incomplete literature search. The numerical tables do show that SPU-MAMBA is competitive with generic PU networks on the synthetic benchmarks, which is a legitimate but narrower result. Since the central claim is about opening a new avenue for single-photon sensing, the correct test is against existing single-photon reconstruction methods, not only generic PU networks. This concern is routed through experimental completeness and the internal logic of the paper, not through circularity or ad hominem. The paper has independent support in that the tables are internally consistent, the method is clearly described, and the architecture choices are well motivated, but the lack of a single-photon-specific comparison set and the unvalidated noise model make the practical utility and novelty claims conditional.","tokens_in":13103,"tokens_out":1450,"duration_ms":14764,"concrete_test":"Re-run the evaluation including at least one single-photon-specific reconstruction baseline from the paper's own Sec. II.B (e.g., first-photon imaging or a Bayesian deconvolution method) on the same noise-augmented PU-GAN and PU1K inputs, computing the same CD, HD, and P2F metrics. Also provide a quantitative evaluation on the real-world single-photon data, e.g., by comparing depth accuracy against a rasterized reference or a reconstructed ground-truth surface, instead of relying solely on visual inspection.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim has two parts: (1) SPU-MAMBA achieves strong reconstruction accuracy and noise robustness on single-photon-like point clouds, and (2) it is the first upsampling framework for single-photon sensing, thereby opening a new avenue for practical applications. Part (1) is supported only by experiments on PU-GAN and PU1K with synthetic Gaussian depth offsets, plus a qualitative real-world comparison with no ground truth. Part (2), the novelty claim, is undercut by the paper's own Sec. II.B, which reviews model-based and learning-based single-photon reconstruction algorithms (e.g., first-photon imaging, Bayesian deconvolution, non-local neural networks, U-net++). None of these single-photon-specific methods is included as a baseline, so the claim of being 'first' is asserted rather than demonstrated. Additionally, the synthetic noise model in Sec. IV is unvalidated: it adds a zero-mean Gaussian depth offset with sigma up to 0.1 to every point, but real single-photon depth bias is systematic and scene-dependent (due to photon arrival statistics, incidence angle, and atmospheric scattering), not necessarily zero-mean Gaussian. This weakens the internal consistency of the robustness claim. The omission is load-bearing because the paper's stated contribution is to the single-photon sensing domain, not merely to generic point upsampling.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SPU-MAMBA, a point upsampling network for single-photon point clouds, built on a Mamba/state-space backbone with a multi-path scanning mechanism, a bidirectional Mamba decoder, and an adaptive upsample shift module. The authors evaluate SPU-MAMBA on the PU-GAN and PU1K benchmarks under synthetic Gaussian depth-offset noise at four levels (σ = 0, 0.02, 0.06, 0.1), comparing with seven generic point upsampling baselines using Chamfer distance, Hausdorff distance, and point-to-surface distance. A qualitative real-world single-photon LiDAR experiment is also reported. The paper claims best or second-best performance across most metrics and noise levels, and states that it is the first to establish the upsampling framework for single-photon sensing.","tokens_in":13367,"tokens_out":7150,"duration_ms":77689,"significance":"If the robustness claims hold, the work would be a useful application of point upsampling to a domain where point clouds are sparse and distorted, and the architectural components are assembled coherently. The paper's strength is its systematic comparison across eight methods and four noise levels on two benchmarks, with internally consistent tables. However, the significance is currently limited by three load-bearing issues: the synthetic Gaussian noise model is not validated against real single-photon distortion, the real-world evaluation is only qualitative, and the point-to-surface metric as defined in Eq. (13) does not measure distance to any ground-truth surface. The absence of single-photon-specific baselines further weakens the domain-novelty claim.","major_comments":[{"comment":"The noise model is unvalidated. Section IV states that a random depth offset sampled from a zero-mean Gaussian with σ ∈ [0,0.1] is added to every point 'to account for the correlation between photon arrival time and emission angle, and the atmospheric scattering effect.' Real single-photon depth errors are not generally zero-mean Gaussian; they include systematic, scene-dependent biases from photon arrival statistics, incidence angle, and atmospheric scattering. The real-world evaluation in Sec. V.D is purely qualitative, so the paper does not demonstrate that the synthetic corruption reproduces the structure of real single-photon point-cloud distortion. Please validate the noise model against real data or provide a quantitative mapping of σ to physical units and real-system parameters.","section":"Sec. IV / Sec. V.D"},{"comment":"The reported performance differences are not accompanied by error bars, confidence intervals, or significance tests. Some claimed improvements are very small; for example, in Table II at σ = 0, SPU-MAMBA's CD is 0.485 × 10⁻³, which is worse than PU-GCN's 0.479 × 10⁻³, and at σ = 0.06 and 0.1 the CD gaps are only 1.8% and 6.5%. Without repeated runs and variance estimates, the central claims of 'best/second-best' performance and 'stronger noise robustness' are not statistically supported.","section":"Sec. V.A / Tables I-II"},{"comment":"The definition of P2F is ambiguous. The text identifies the point-to-surface distance P2F with Eq. (13), but Eq. (13) as written measures the distance between deformed kernel points and the input neighborhood N(p), not the distance to any ground-truth surface. A model that keeps output points near the input can minimize this quantity without improving reconstruction fidelity. If the reported P2F is instead computed against a ground-truth mesh or surface, that metric must be defined separately and reported; as written, the P2F improvements in Tables I and II are not a meaningful fidelity claim.","section":"Eq. (13) / Tables I-II"},{"comment":"The novelty claim is not backed by the comparison set. The abstract and introduction state that this is the first upsampling framework for single-photon sensing, but all baselines are generic point upsampling networks; none of the single-photon-specific reconstruction methods reviewed in Sec. II.B (e.g., first-photon imaging [1], non-local neural network [40], U-Net++ [41], or pixel-wise residual shrinkage net [42]) is included in the experiments. To substantiate the domain-specific contribution, at least one single-photon reconstruction or denoising method should be adapted as a baseline, or the novelty claim should be tempered.","section":"Sec. II.B / Sec. V"}],"minor_comments":[{"comment":"The symbol σ is used both for the noise standard deviation in Sec. IV and for the distance scaling factor in Eqs. (13) and (14); please use different symbols to avoid ambiguity.","section":"Eqs. (13)-(14) / Sec. IV"},{"comment":"Please state explicitly that the synthetic noise is applied only to the input point clouds and not to the ground-truth point clouds; the text 'added to every point in the dataset' is ambiguous.","section":"Sec. V.C"},{"comment":"Appendix A is referenced in Sec. II.C but does not appear in the manuscript; please include it or remove the reference.","section":"Sec. II.C"},{"comment":"In the Conclusions, 'sing-photon point cloud' should be corrected to 'single-photon point cloud'.","section":"Sec. VII"},{"comment":"The claim of linear computational complexity should be qualified: the multi-path scanning and the K-nearest-neighbor construction in Sec. III.C involve sorting and neighbor search, which are not strictly linear in the number of points.","section":"Sec. III"},{"comment":"The real-world evaluation would be strengthened by a quantitative assessment of uniformity or detail preservation, since the current visual comparison is only qualitative.","section":"Sec. V.D / Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable engineering contribution, but the domain-specific validation is the main weakness. The unvalidated Gaussian noise model and the ambiguous P2F metric are load-bearing for the central claims; both are fixable with additional experiments and clarifications. The 'first to establish the upsampling framework' claim should be either supported with a more targeted baseline comparison or softened. I would not reject the paper outright, but the authors need to address these points before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper does something sensible and reports it cleanly, but it overclaims the single-photon part. The core idea—using a point cloud upsampling network to densify and de-distort SPAD-derived point clouds—is new, as far as I can tell, and the SPU-MAMBA architecture (bidirectional Mamba with multi-path scanning, deformable KPConv shift) is a reasonable composition of known components. The experiments are thorough for the PU literature: two benchmarks, four noise levels, seven baselines, plus a qualitative real-world example. The tables are internally consistent and the method is best or near-best on most metrics. That is a genuine empirical result for upsampling on Gaussian-corrupted point clouds.\n\nThe soft spots are in the single-photon framing. The synthetic noise model adds a zero-mean Gaussian depth offset to every point. Real single-photon depth bias is systematic, scene-dependent, and often not zero-mean; the paper asserts the Gaussian proxy captures the relevant physics but does not validate that claim. Without validation, the robustness numbers say little about real single-photon sensing. Relatedly, the 'first to establish the upsampling framework' claim is undercut by the absence of any single-photon-specific reconstruction or super-resolution baseline. Those methods do solve a different problem (reconstructing from histograms), so the omission is not fatal to the architecture story, but it weakens the practical-utility argument.\n\nMinor but worth noting: no error bars or significance tests, so small gaps (e.g., 1.8% in CD) could be noise; the real-world evaluation is visual only; and no code or data is released, which makes the numbers hard to verify.\n\nWho is this for? Point cloud upsampling people will find the architecture interesting. Single-photon imaging people will, rightly, be skeptical of the transfer claim. It deserves a serious referee, but the authors need to either validate the noise model against real single-photon data or soften the claims, and ideally add at least one single-photon-specific baseline. I'd send it to review, not desk-reject, but with a request for those revisions and error bars.\n\nBest.","headline":"A solid point-upsampling paper with a new application to single-photon clouds, but the single-photon relevance rests on an unvalidated Gaussian noise model and the 'first' claim is asserted, not shown.","tokens_in":13935,"tokens_out":3117,"would_cite":false,"duration_ms":31305,"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":"The paper proposes treating single-photon point cloud reconstruction as a point upsampling task, and reports that a Mamba-based network, SPU-MAMBA, achieves best or near-best accuracy under distortion noise.","keywords":["single-photon sensing","point cloud upsampling","state space model","Mamba","depth distortion","LiDAR","point cloud densification","noise robustness"],"falsifier":"Evaluate SPU-MAMBA and the baseline methods on real single-photon LiDAR scenes with known ground-truth geometry (for example, a calibrated target scene scanned at multiple ranges), measure the same CD, HD, and P2F metrics against the ground truth, and check whether SPU-MAMBA still holds its ranking. As a quicker check, measure the actual distribution of depth errors in a real SPAD point cloud: if the error is non-Gaussian, has a non-zero mean, or is strongly correlated with surface reflectivity or incidence angle, then the paper's noise-augmentation procedure does not capture the true distortion and the reported robustness may not hold outside the synthetic setup.","tokens_in":1966,"feed_emoji":"📡","tokens_out":2209,"duration_ms":62023,"temperature":0.7,"pith_summary":"Single-photon sensing yields point clouds that are sparse and spatially biased, which limits their practical use in 3D tasks. This paper argues that the problem can be framed as point cloud upsampling: a neural network takes the sparse, distorted cloud and generates additional points on the underlying surface while correcting depth offset. The proposed network, SPU-MAMBA, uses a state-space (Mamba) backbone with multi-path scanning and an adaptive shift module, achieving linear computational complexity. On standard benchmarks with added depth-offset noise, it reports the best or second-best Chamfer distance, Hausdorff distance, and point-to-surface error among eight methods, and shows flatter degradation as noise increases. If correct, this offers a direct post-processing route to make single-photon imaging outputs denser and more geometrically faithful for downstream tasks.","feed_headline":"A Mamba-based network densifies and corrects single-photon point clouds","feed_subtitle":"Upsampling with a state-space model beats eight baselines on distorted 3D point clouds.","key_machinery":"The load-bearing machinery is the SPU-MAMBA network architecture, organized into a point encoder, a Mamba decoder, and an upsample shift module. The decoder's multi-path scanning mechanism (six axis-ordered sortings such as XYZ, XZY, YXZ, and their permutations) serializes a point cloud into permutation-equivariant sequences, preserving spatial coherence and depth-layer continuity. Bidirectional Mamba blocks — a state-space model with forward and backward passes plus convolutional local-feature extraction — process each sequence in linear time, suppressing noise propagation while capturing long-range geometry. The upsample shift module uses a two-stage deformable kernel-point convolution to predict per-point displacements for $r_c$ generated points, with a repulsion loss that encourages uniform distribution. The key combination is that the scanning order captures long-range geometry, the bidirectional SSM balances global and local detail, and the deformable shift corrects offset-induced distortion, together producing the reported accuracy and robustness.","core_discovery":"The central claim is that point upsampling (PU) networks can serve as a general post-processing tool for single-photon sensing: given a sparse, distorted point cloud, a PU network can increase point density by a factor $r_c$ and simultaneously reduce offset-induced distortion, yielding a more uniform and geometrically accurate representation. The specific instantiation, SPU-MAMBA, integrates a multi-path scanning mechanism that serializes points along six axis-ordered paths, a bidirectional Mamba (state-space) backbone that captures global context and local detail with linear complexity, and an upsample shift module based on deformable kernel-point convolution that displaces generated points to fit local geometry. On the PU-GAN and PU1K benchmarks with additive Gaussian depth offsets of $\\sigma$ up to 0.1, the authors report that SPU-MAMBA scores the best Chamfer distance at most noise levels, consistently second-best Hausdorff distance, and top or near-top point-to-surface error, while degrading more gracefully than baselines. On real-world single-photon LiDAR data, it is claimed to generate visually consistent, detail-preserving, and noise-suppressed point clouds.","pith_inferences":["If the synthetic zero-mean Gaussian depth-offset model matches the real bias structure, the same network could be trained entirely on simulated LiDAR histograms, since the loss functions only require the noisy input and a clean surface model — no paired real-world ground truth needed.","The multi-path scanning mechanism may transfer to other modalities that produce sparse, distorted point clouds, such as underwater sonar or long-range radar, where similar reflectivity-driven sampling bias occurs.","A testable extension is to quantify whether the upsampled point clouds actually improve downstream task accuracy (for example, object detection mAP or segmentation IoU), which the paper motivates but does not measure directly.","The robustness claim is demonstrated only for zero-mean Gaussian depth offsets; real distortions are likely correlated with reflectivity and incidence angle, so a stress test with a physics-based noise model would clarify the practical limits of the approach."],"forward_implications":["SPU-MAMBA can upsample non-uniformly distributed point clouds with linear computational complexity, making the approach scalable to large-scale scenes where quadratic attention would be prohibitive.","A single model, when trained on noise-augmented point clouds, maintains more stable Chamfer distance and point-to-surface error as depth-offset noise increases from $\\sigma=0$ to $\\sigma=0.1$, degrading more gracefully than the eight baselines.","The framework can be applied directly to real-world single-photon LiDAR data without ground truth, producing point clouds that are visually more uniform, detail-preserving, and resilient to residual noise and outliers.","Because the upsampled clouds better preserve local geometry and reduce spatial distortion, downstream tasks such as object detection, segmentation, and geometric classification stand to receive higher-quality input point clouds.","The authors position this as the first upsampling framework for single-photon sensing, opening a distinct processing stage that complements model-based denoising and learning-based reconstruction."],"supporting_citations":[{"why":"Supplies the Mamba state-space architecture that the SPU-MAMBA backbone is built on, providing selective scanning and linear-time sequence modeling.","marker":"[31]"},{"why":"Defines the point cloud upsampling task with PU-Net, one of the eight baseline methods and the origin of the dense point generation objective.","marker":"[34]"},{"why":"Provides the PU-GAN dataset and adversarial baseline, the first of the two main benchmarks used for evaluation.","marker":"[48]"},{"why":"Introduces the PU1K dataset and the PU-GCN graph-convolutional baseline, the second main benchmark and one of the strongest comparison methods.","marker":"[35]"},{"why":"Grad-PU, a baseline using gradient descent with learned distance functions; its spherical uniform sampling with fixed radius is the approach the upsample shift module is designed to improve upon.","marker":"[37]"},{"why":"Provides the deformable kernel-point convolution (KPConv) that the upsample shift module uses to predict displacements and preserve sharp features.","marker":"[47]"},{"why":"Point Mamba, a state-space backbone for point clouds with octree-based ordering, which the multi-path scanning mechanism extends with six axis-ordered paths.","marker":"[33]"},{"why":"Supplies the photon flux model that links photon arrival time to depth, the basis for the Gaussian depth-offset noise augmentation and the point cloud generation from histograms.","marker":"[39]"}],"fun_headline_variants":["First upsampling network for single-photon sensing","SPU-MAMBA: densifies and corrects photon point clouds","State-space model upsamples sparse single-photon point clouds","Bidirectional Mamba backbone densifies single-photon point clouds","Upsampling with Mamba corrects distortion in single-photon point clouds"],"cache_read_input_tokens":16000,"weakest_assumption_plain":"The synthetic zero-mean Gaussian depth offset added to every point is assumed to reproduce the real bias structure of single-photon point clouds, which actually arises from the correlation between photon arrival time and emission angle and from atmospheric scattering; if it does not, the reported robustness gains will not transfer to real single-photon sensing.","fun_headline_variants_meta":{"raw":{"variants":["First upsampling network for single-photon sensing","SPU-MAMBA: densifies and corrects photon point clouds","State-space model upsamples sparse single-photon point clouds","Bidirectional Mamba backbone densifies single-photon point clouds","Upsampling with Mamba corrects distortion in single-photon point clouds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000712,"raw_usage":{"total_tokens":3207,"prompt_tokens":951,"completion_tokens":2256,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":2168}},"tokens_in":567,"tokens_out":2256,"duration_ms":16832,"temperature":1.0,"reasoning_tokens":2168,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:16:48.055717+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate SPU-MAMBA and the baseline methods on real single-photon LiDAR scenes with known ground-truth geometry (for example, a calibrated target scene scanned at multiple ranges), measure the same CD, HD, and P2F metrics against the ground truth, and check whether SPU-MAMBA still holds its ranking. As a quicker check, measure the actual distribution of depth errors in a real SPAD point cloud: if the error is non-Gaussian, has a non-zero mean, or is strongly correlated with surface reflectivity or incidence angle, then the paper's noise-augmentation procedure does not capture the true distortion and the reported robustness may not hold outside the synthetic setup.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the point cloud upsampling task with PU-Net, one of the eight baseline methods and the origin of the dense point generation objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PU-GAN dataset and adversarial baseline, the first of the two main benchmarks used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the PU1K dataset and the PU-GCN graph-convolutional baseline, the second main benchmark and one of the strongest comparison methods."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Grad-PU, a baseline using gradient descent with learned distance functions; its spherical uniform sampling with fixed radius is the approach the upsample shift module is designed to improve upon."},{"cited_title":"Thomas, C","cited_arxiv_id":null,"evidence_quote":"Provides the deformable kernel-point convolution (KPConv) that the upsample shift module uses to predict displacements and preserve sharp features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the photon flux model that links photon arrival time to depth, the basis for the Gaussian depth-offset noise augmentation and the point cloud generation from histograms."}],"review_version":2}