{"id":"7d40f1e2-eafe-47f9-965c-a625279e6e11","arxiv_id":"2506.15806","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A neural network learns obstacle shapes as Signed Distance Functions from LiDAR, but only training-loss results are reported and the collision-safety claim is not tested.","lead":"The paper trains a small neural network on LiDAR point clouds to build a signed distance field, a continuous 3D map of how far every point is from the nearest obstacle. The authors claim this gives self-driving cars better obstacle boundaries than bounding boxes, but the experiments report training losses only and never test collision behavior.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption that fabricated negative SDF samples beyond LiDAR termination points encode true occupancy is untested and likely false for real scenes, so the learned zero-level set may not match obstacle geometry and the collision-detection claim is unsupported.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing weakness: fabricated negative SDF distances sampled beyond LiDAR termination points may not correspond to true occupied space. I agree with that identification, and it is the single most important vulnerability because it undermines the only source of inside-object supervision. The paper's own results contain direct symptoms: invalid negative confidence outputs in Section VII-A, no validation against ground-truth geometry, and no collision or reconstruction experiments. The central claim in the abstract is therefore not supported by the evidence presented. I do not see a separate concern that is more load-bearing than this one, and I do not think the concern is merely a disagreement with consensus; it is an internal gap between the label-generation procedure and the geometric meaning of an SDF. The proposed test would settle the question directly by measuring how often the generated negative labels are geometrically wrong and whether fixing them changes the learned surface. Since the reader already recommends rejection and this concern supports that outcome, no verdict adjustment is needed.","tokens_in":6591,"tokens_out":4849,"duration_ms":64592,"concrete_test":"Take one annotated nuScenes vehicle with a fitted watertight CAD mesh (or generate a synthetic vehicle in CARLA/Blender), raycast from a sensor pose, and apply the paper's uniform and Gaussian augmentation procedures exactly as described in Section VI-A. For each generated negative sample point, compute its true sign by checking whether it lies inside the watertight mesh, and report the fraction of negative labels that are actually in free space. Then train the proposed ANN on the original labels and on the corrected labels, and compare the predicted zero-level sets against the mesh using Chamfer distance and volumetric IoU. If a non-negligible fraction (e.g., >5%) of negative labels are mislabeled, or if the corrected labels materially change the reconstructed surface, the current loss numbers cannot support the collision-detection claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section VI-A states that negative points were generated beyond the LiDAR termination point and then truncated. A termination point is the first surface hit along a ray; beyond that point the space is not necessarily occupied. For a thin or non-watertight object the region past the back face is free space, and for a near object the same ray can pass through a gap before hitting another object. The method has no back-surface, mesh, or multi-view occupancy information, so the sign and magnitude of these negative labels are fabrications. This matters because negative samples are the only source of inside-object supervision, and the reported loss reductions (e.g., 0.37 to 0.18 in Table I) only measure fit to this self-generated label set. Section VII-A also reports invalid confidence values for negative SDF predictions, and no reconstruction or collision benchmark is provided. If the fabricated negative labels are wrong, the zero-level set does not represent obstacle boundaries, and the central claim about enhancing collision detection collapses regardless of model architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a learning-based method for implicit 3D scene reconstruction of static obstacles from LiDAR point clouds, using a fully connected network with Fourier feature encoding to predict a signed distance function (SDF) and a per-point confidence. The authors evaluate three design choices: uniform versus Gaussian augmentation along LiDAR rays, model depth with and without skip connections, and the addition of a Fourier feature encoder. The only reported quantitative results are Huber losses on the authors' own training setup, and the paper also reports invalid confidence predictions for negative SDF values. The central claim is that the resulting boundary-level SDF representation would enhance collision detection in congested driving scenarios, but no collision-detection experiment or reconstruction metric against ground truth is provided.","tokens_in":6818,"tokens_out":3626,"duration_ms":42904,"significance":"If the claimed result held, a compact implicit SDF representation with boundary-level accuracy could be a useful alternative to bounding-box representations for proximity estimation in dense traffic. The work addresses a relevant problem and includes useful ablations of augmentation strategies, Fourier features, and model capacity. However, the paper's evidence is far from sufficient: the only numerical results are training losses on self-generated labels, there is no held-out evaluation, no reconstruction metric, no comparison with baselines, and no collision experiment. The paper also explicitly reports confidence values outside the allowed probability range. The direction is interesting, but the current manuscript does not support its central claims.","major_comments":[{"comment":"The negative SDF labels are fabricated by generating points beyond the LiDAR termination point and truncating them. A termination point is the first surface hit along a ray, so the region beyond it is not necessarily occupied; for thin or non-watertight objects, or through gaps, it can be free space. Since these negative samples are the only source of inside-object supervision, the learned zero-level set may not correspond to actual obstacle boundaries. The paper never validates the sign or magnitude of these labels against true occupancy (e.g., meshes, multi-sweep consistency, or manual labels), and the collision-detection claim depends directly on this assumption.","section":"Section VI-A"},{"comment":"The paper reports that for negative SDF values the model produces invalid confidence scores, including values from 0 to -3 for the uniform augmentation and 0 to -0.5 for the Gaussian augmentation. Equation (1) is described as producing a probability, but the reported outputs are not probabilities, and the range of the hyperparameter b is never specified. Without a meaningful and calibrated confidence output, the proposed confidence mechanism cannot support any collision-reasoning application.","section":"Section VII-A and Eq. (1)"},{"comment":"The only quantitative results are Huber losses computed on the authors' own training setup. There is no held-out validation or test set, no error bars, no comparison against ground-truth SDF values (e.g., Chamfer distance, F-score, voxel IoU), and no comparison with a baseline representation such as bounding boxes. The reported loss reductions (from 0.37 to 0.18) measure fit to self-generated labels, not reconstruction accuracy or collision-relevant performance.","section":"Table I and Section VI-B"},{"comment":"The central claim that the method 'would significantly enhance collision detection performance' is not tested anywhere in the manuscript. No collision-detection, proximity-estimation, or planning experiment is reported, and no metric such as distance-to-obstacle error in dense traffic is computed. The claim is supported only by training-loss comparisons and must either be substantiated with appropriate experiments or removed.","section":"Abstract and Section IX"}],"minor_comments":[{"comment":"The sentence describing the network configuration is ambiguous: 'with tanh, Huber loss, Adam, 0.4 as activation function, loss function, optimizer, and learning rate' does not clearly map each term to its role, and a learning rate of 0.4 would be unusually high and should be justified.","section":"Section IV"},{"comment":"The phrase 'output two values, the sine distance along with a confidence value' contains a typo; it should read 'signed distance'.","section":"Section IV"},{"comment":"The Directed Hausdorff distance threshold used for scene selection is not reported or justified, and the ground-plane threshold of -1.563 meters is presented without context.","section":"Section V"},{"comment":"The bullet list under 'Requirements when calculating confidence score' is difficult to parse, and the text preceding Eq. (1) does not clearly define d(p) and d_max in relation to each LiDAR ray.","section":"Section III"},{"comment":"There are two subsections titled 'Limitations of point sampling using LIDAR sensor' (B and C), which is confusing; the second should be retitled.","section":"Section VIII"},{"comment":"The manuscript contains numerous grammatical errors and typos, such as 'build a the static Signed Distance Function (SDF) maps', 'the sign distance function represents any shape through parameters', and 'Object maps ... It depicts actual sceneries'. A careful proofreading pass is needed.","section":"Abstract and throughout"}],"recommendation":"reject","confidential_remarks":"The paper's central claim is unsupported by the reported experiments, and the negative-label generation assumption in Section VI-A is a fundamental, unvalidated step. A revision that replaces the fabricated negative samples with actual occupancy supervision and adds held-out reconstruction and collision benchmarks would be a substantially different contribution. I therefore recommend rejection in the current form, though the underlying problem is worth pursuing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a pilot study with one genuinely useful ablation and a load-bearing evaluation gap. The Fourier-feature comparison (0.37 to 0.18 Huber loss in Table I) and the finding that bigger networks do not help are real observations about training dynamics for LiDAR-based SDF fields. The paper also honestly reports its own negative results: invalid confidence values in Section VII-A and no improvement from depth. That is more candid than many submissions.\n\nWhat is actually new: the uniform-versus-Gaussian comparison for negative sampling along LiDAR rays, and the specific finding that Gaussian sampling near the termination point plus a Fourier encoder halves the loss. That is a modest but legitimate methodological result for anyone building SDF maps from sparse LiDAR.\n\nThe soft spots are serious. The abstract claims the method 'would significantly enhance collision detection performance,' but there is no collision experiment, no reconstruction metric against ground truth, no held-out validation, and no error bars. The only quantitative evidence is training loss on self-generated labels. More fundamentally, the negative SDF labels are fabricated: Section VI-A generates points beyond the LiDAR termination point and treats them as inside the obstacle. Beyond the first surface hit, the space is not necessarily occupied—for thin objects, the ray can pass through to free space, or hit a second object further along. The authors never verify occupancy with a mesh or multi-view data, so the zero-level set may not correspond to obstacle boundaries. This is load-bearing, not cosmetic: if the negative labels are wrong, the collision-avoidance claim collapses regardless of architecture.\n\nThe novelty framing also does not hold up. The paper calls itself 'the first attempt in the autonomous driving research community' while citing MV-DeepSDF [9], Autolabeling [8], and Deep SDF Maps [11], all of which apply SDF learning in driving or navigation. The contribution is an application with a specific sampling comparison, not a new class of method.\n\nThe appended limitations (Section VIII) discuss distance-dependent noise and small objects, but they do not flag the fabricated-negative-label issue, which is the one that matters.\n\nBottom line: this is a useful workshop-level ablation study in search of a proper evaluation. A serious referee could help the authors add real reconstruction and collision metrics, but as submitted the central claim is unsupported. If it comes to you, I would send it back for major revision or reject with clear guidance.","headline":"Useful pilot ablation, but the load-bearing SDF labels are fabricated from LiDAR ray geometry and no collision experiment exists, so the main claim is unsupported.","tokens_in":7328,"tokens_out":2908,"would_cite":false,"duration_ms":33242,"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":"Sparse LiDAR scans can be turned into continuous signed-distance obstacle maps by a small neural net.","keywords":["signed distance function","implicit 3D reconstruction","LiDAR point clouds","autonomous driving","collision avoidance","Fourier feature encoding","NuScenes dataset","neural network"],"falsifier":"Reconstruct a NuScenes scene with this pipeline and compare the predicted zero-level surface $\\{p:S(p)=0\\}$ against a ground-truth watertight mesh or dense multi-sweep occupancy map of the same scene; if the synthetic negative points are not actually inside occupied space, the surface will be systematically offset and interior/exterior queries will disagree with true occupancy.","tokens_in":6394,"feed_emoji":"🚗","tokens_out":5449,"duration_ms":60054,"temperature":0.7,"pith_summary":"This paper tries to establish that a neural network can learn a static Signed Distance Function (SDF) from LiDAR point clouds, representing surrounding obstacles with boundary-level detail in a form that suits collision avoidance in congested traffic better than the bounding-box or polygonal maps now used. The proposed pipeline filters background points from NuScenes scenes, augments each LiDAR ray with synthetic interior and exterior samples, labels them by nearest-surface distance, and trains a small fully connected network with Fourier feature encoding to output signed distance and a confidence value. The direct evidence reported is reconstruction-oriented: adding a Fourier feature encoder reduces the Huber loss from 0.4 to 0.18, and Gaussian negative sampling near the LiDAR termination point yields fewer invalid confidence values than uniform sampling. No collision-detection experiment is reported, so the safety benefit is an extrapolation from reconstruction accuracy rather than a measured outcome.","feed_headline":"LiDAR point clouds become continuous collision maps","feed_subtitle":"A small neural net learns signed-distance obstacle surfaces from sparse scans, promising safer dense-traffic navigation.","key_machinery":"The central object is the Signed Distance Function $S(p)$, the distance from a query point $p\\in\\mathbb{R}^3$ to the nearest obstacle surface, with negative sign inside the watertight boundary. A small fully connected network with a Fourier feature encoder maps spatial inputs to a predicted $(S(p), C(p))$ pair, where the zero level set $\\{p:S(p)=0\\}$ is taken as the obstacle shape and $C(p)$ is a confidence value computed by a formula borrowed from prior work that downweights large negative distances. The K-D tree accelerates the nearest-neighbour distance labelling needed to build training targets.","core_discovery":"The central claim is that implicit 3D reconstruction through a learned SDF gives a continuous, storage-efficient map of obstacle surfaces, and that the zero set of the network defines the obstacle boundary more accurately than traditional polygonal representations. On NuScenes static scenes, a three-layer fully connected network with $\\tanh$ activation, Huber loss, Adam, and Fourier feature encoding fits SDF labels, and the reported experiments show the Fourier encoder roughly halves the loss (from 0.4 to 0.18) and keeps invalid confidence predictions for interior points within $0$ to $-0.5$, compared with $0$ to $-3$ under uniform sampling. The paper presents this as the first learning-based SDF reconstruction aimed specifically at collision understanding in autonomous driving, and argues that boundary-level detail is precisely what dense, dynamic traffic requires.","pith_inferences":["Editorial: the synthetic negative points beyond the LiDAR termination point assume that every obstacle is solid behind its visible surface; a direct test is to compare the learned field against dense multi-sweep occupancy or watertight CAD models of the same scenes.","Editorial: the same SDF field could feed motion planning directly through its gradient as a repulsive potential, turning reconstruction fidelity into a quantitative safety margin.","Editorial: extending the static field to dynamic obstacles would require conditioning the network on time or object identity, and the confidence output could then double as an occupancy-uncertainty measure for occluded regions."],"forward_implications":["An accurate learned SDF lets a vehicle query the distance from any point to any obstacle surface directly, so proximity checks no longer depend on stored meshes or bounding-box approximations.","The confidence output can flag unreliable interior distances, giving downstream planners a principled way to distrust predictions in occluded or far-away regions.","Fourier feature encoding is the main accuracy lever in the reported tests, so future LiDAR-based SDF pipelines should include it.","Shallower networks outperform deeper ones beyond roughly 18 layers, which keeps online reconstruction costs low.","Gaussian negative sampling near the LiDAR termination point is preferable to uniform sampling because it reduces invalid confidence scores for negative distances."],"supporting_citations":[{"why":"Supplies the LiDAR point clouds and class-filtered static scenes used for all reconstruction experiments.","marker":"[12]"},{"why":"Provides the confidence-calculation formula that downweights unreliable negative signed distances.","marker":"[11]"},{"why":"Introduces the DeepSDF-based auto-labelling and differentiable rendering line of implicit reconstruction that motivates the approach.","marker":"[8]"},{"why":"Demonstrates multi-sweep SDF vehicle reconstruction in autonomous driving, the direct precursor the paper extends.","marker":"[9]"},{"why":"Shows SDF-2-SDF registration for real-time reconstruction, supporting the use of SDF as a continuous reconstruction representation.","marker":"[7]"}],"fun_headline_variants":["Neural SDF from LiDAR sharpens collision boundaries","Deep SDF maps obstacles for safer dense traffic","Implicit 3D reconstruction learns obstacle surfaces from scans","Learned SDF maps from LiDAR boost collision safety","Implicit SDF from LiDAR enhances collision detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that points generated beyond each LiDAR termination point really lie inside an obstacle, so their negative distances are valid ground truth; if they fall in free space, the learned field is artificial and every downstream collision-use claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Neural SDF from LiDAR sharpens collision boundaries","Deep SDF maps obstacles for safer dense traffic","Implicit 3D reconstruction learns obstacle surfaces from scans","Learned SDF maps from LiDAR boost collision safety","Implicit SDF from LiDAR enhances collision detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000728,"raw_usage":{"total_tokens":3248,"prompt_tokens":922,"completion_tokens":2326,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":2248}},"tokens_in":538,"tokens_out":2326,"duration_ms":18107,"temperature":1.0,"reasoning_tokens":2248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:50:48.408254+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct a NuScenes scene with this pipeline and compare the predicted zero-level surface $\\{p:S(p)=0\\}$ against a ground-truth watertight mesh or dense multi-sweep occupancy map of the same scene; if the synthetic negative points are not actually inside occupied space, the surface will be systematically offset and interior/exterior queries will disagree with true occupancy.","supporting_citations":[{"cited_title":"Autolabeling 3D Objects with Differentiable Rendering of SDF Shape Priors","cited_arxiv_id":"1911.11288","evidence_quote":"Introduces the DeepSDF-based auto-labelling and differentiable rendering line of implicit reconstruction that motivates the approach."},{"cited_title":"MV-DeepSDF: Implicit Modeling with Multi-Sweep Point Clouds for 3D Vehicle Reconstruction in Autonomous Driving","cited_arxiv_id":"2309.16715","evidence_quote":"Demonstrates multi-sweep SDF vehicle reconstruction in autonomous driving, the direct precursor the paper extends."},{"cited_title":"SDF-2-SDF Registration for Real-Time 3D Reconstruction from RGB-D Data,","cited_arxiv_id":null,"evidence_quote":"Shows SDF-2-SDF registration for real-time reconstruction, supporting the use of SDF as a continuous reconstruction representation."}],"review_version":1}