{"id":"c19f2569-3cfa-47ca-a0c9-00b72b9f9af9","arxiv_id":"2406.04301","paper_version":4,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"EpiS improves generalizable neural surface reconstruction from sparse views by guiding epipolar feature aggregation with cost volumes, using an epipolar transformer, and applying pretrained monocular depth constraints, outperforming prior methods on DTU and BlendedMVS.","lead":"The paper introduces EpiS, a neural framework for surface reconstruction from sparse multi-view images that aggregates epipolar features guided by cost volumes and adds monocular depth regularization. Smart generalists might read it to see how epipolar geometry can address over-smoothing in practical 3D reconstruction tasks with limited views.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Monocular depth regularization may inject biases misaligned with sparse-view epipolar geometry","rationale":"The identified concern directly matches the reader's weakest assumption. Full-text details on the loss formulation would be needed to confirm whether the scale-invariance is derived in a way that guarantees multi-view consistency, but the assumption remains the least-secured link for the generalization claim.","tokens_in":1700,"tokens_out":299,"duration_ms":20906,"concrete_test":"Ablate the geometry regularization loss on the DTU 3-view split while keeping the epipolar transformer fixed; recompute mean Chamfer distance and normal consistency on the same test scenes. If the metric degrades by >15% relative to the full model (or matches the no-regularization baseline), the regularization does not reliably support the headline outperformance.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the geometry regularization (scale-invariant global/local constraints from a pretrained monocular depth model) supplies unbiased signals that reinforce rather than contradict the epipolar feature aggregation. If the monocular predictions contain systematic errors (e.g., in low-texture or view-dependent regions common under 3-view DTU/BlendedMVS protocols), the ray-wise SDF features will be pulled toward inconsistent surfaces, undermining the reported gains over cost-volume baselines. The abstract presents this as a mitigation strategy without evidence that the constraints remain consistent with multi-view geometry after scale normalization.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes EpiS, a generalizable neural surface reconstruction framework for sparse multi-view inputs. It replaces direct regression from cost-volume statistics with coarse cost-volume features guiding aggregation of fine-grained epipolar features sampled along epipolar lines, fused via an epipolar transformer and ray-wise aggregation to produce SDF-aware features. A geometry regularization strategy adds scale-invariant global and local constraints from a pretrained monocular depth model. Experiments on DTU and BlendedMVS report significant outperformance over prior generalizable methods under sparse views without per-scene optimization.","tokens_in":1801,"tokens_out":560,"duration_ms":20922,"significance":"If the reported gains hold after verification of implementation details and ablations, the explicit use of epipolar geometry for feature aggregation combined with monocular regularization could advance sparse-view surface reconstruction by preserving view-dependent structure that simple cost-volume statistics discard.","major_comments":[{"comment":"Abstract (geometry regularization strategy paragraph): The central performance claim depends on the monocular depth constraints supplying unbiased signals that align with multi-view epipolar geometry after scale normalization. No analysis or test is described showing that systematic errors in the pretrained model (e.g., in low-texture or view-dependent regions under 3-view DTU/BlendedMVS protocols) do not pull ray-wise SDF features toward inconsistent surfaces, which directly risks undermining the reported gains over cost-volume baselines.","section":"Abstract"},{"comment":"Method description (epipolar feature aggregation): The claim that coarse cost-volume features reliably guide fine-grained epipolar aggregation is load-bearing for the outperformance result, yet the manuscript provides no quantitative measure (e.g., alignment error or ablation removing the guidance) of how well this guidance functions when the cost volume itself is severely under-constrained by only three views.","section":"Method"}],"minor_comments":[{"comment":"The abstract and method sections use 'SDF-aware features' without an explicit definition or equation linking the ray-wise aggregation output to the signed distance function used for surface extraction.","section":"Abstract"},{"comment":"Dataset splits, number of views (e.g., exact 3-view protocol), and whether error bars or multiple runs are reported are not mentioned in the provided abstract; these details are needed for reproducibility of the 'significantly outperforms' claim.","section":"Experiments"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to rely on a specific pretrained monocular depth model whose choice could affect results; confirming whether this model and its training data are fully disclosed would be useful for assessing potential data leakage or bias alignment with the evaluation sets."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback highlighting two important aspects of our method that warrant further clarification. We address each major comment below and indicate where revisions will be made.","responses":[{"response":"We agree that an explicit analysis of potential systematic biases in the pretrained monocular depth model under the 3-view protocols would strengthen the paper. The scale-invariant global and local constraints are intended to reduce sensitivity to absolute scale and local inconsistencies, and the reported gains over pure cost-volume baselines on both DTU and BlendedMVS provide indirect evidence that any residual biases do not dominate. Nevertheless, we will add a dedicated paragraph in the revised manuscript discussing known limitations of monocular depth estimators in low-texture and view-dependent regions, together with qualitative visualizations of the depth predictions used during training on the evaluation scenes.","revision_made":"partial","referee_comment":"[Abstract] Abstract (geometry regularization strategy paragraph): The central performance claim depends on the monocular depth constraints supplying unbiased signals that align with multi-view epipolar geometry after scale normalization. No analysis or test is described showing that systematic errors in the pretrained model (e.g., in low-texture or view-dependent regions under 3-view DTU/BlendedMVS protocols) do not pull ray-wise SDF features toward inconsistent surfaces, which directly risks undermining the reported gains over cost-volume baselines."},{"response":"The guidance mechanism is indeed central. While the current manuscript does not report a direct alignment-error metric between coarse cost-volume features and the sampled epipolar features, the ablation studies already isolate the contribution of the epipolar transformer and ray-wise aggregation. To directly quantify the guidance quality under three-view sparsity, we will add a new ablation that replaces the learned guidance with uniform or random sampling along epipolar lines and report the resulting surface reconstruction metrics on DTU.","revision_made":"yes","referee_comment":"[Method] Method description (epipolar feature aggregation): The claim that coarse cost-volume features reliably guide fine-grained epipolar aggregation is load-bearing for the outperformance result, yet the manuscript provides no quantitative measure (e.g., alignment error or ablation removing the guidance) of how well this guidance functions when the cost volume itself is severely under-constrained by only three views."}],"tokens_in":1380,"tokens_out":486,"duration_ms":14127,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing here is EpiS, which samples fine epipolar features along lines in source views, guided by coarse cost-volume features, then fuses them with an epipolar transformer and adds scale-invariant constraints from a pretrained monocular depth model. This is meant to keep more geometric structure than the mean/variance summaries common in cost-volume baselines for generalizable surface reconstruction from sparse inputs like three views on DTU or BlendedMVS.","headline":"EpiS uses cost-volume-guided epipolar sampling plus monocular depth reg to target over-smoothing in sparse-view SDF recon, but the monocular step is the part that needs checking.","tokens_in":2272,"tokens_out":171,"would_cite":false,"duration_ms":18021,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[{"relation":"unclear","rs_module":"IndisputableMonolith/Cost/FunctionalEquation.lean","rs_theorem":"washburn_uniqueness_aczel","paper_passage":"EpiS uses coarse cost-volume features to guide the aggregation of fine-grained epipolar features sampled along corresponding epipolar lines... epipolar transformer fuses multi-view information, followed by ray-wise aggregation to produce SDF-aware features... geometry regularization strategy that leverages a pretrained monocular depth model through scale-invariant global and local constraints (global triplet loss, local gradient loss)."},{"relation":"unclear","rs_module":"IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean","rs_theorem":"absolute_floor_iff_bare_distinguishability","paper_passage":"Epipolar & Ray Information Aggregation... Linearized Attention mechanism... Geometry Decoder & Weights Decoder... Lglobal = ((d̂1 − d̂s) × (d̃2 − d̃s) − (d̂2 − d̂s) × (d̃1 − d̃s))², Llocal = (1 − v̂ · ṽ / ||v̂||·||ṽ||)²"}],"headline":"Epipolar transformer + monocular triplet/gradient regularization unrelated to RS J-cost or φ-ladder","alignment":"orthogonal","rationale":"Paper centers on cost-volume-guided epipolar attention (linearized cross-attention along rays) plus scale-invariant triplet + derivative losses from a pretrained monocular depth model to produce SDF features. No ratio-symmetric cost, no J(x) = ½(x + x⁻¹) − 1, no cosh(ρ ln φ) form, no 8-tick or φ-ladder structure, no parameter-free constant derivation. Domain is practical CV surface reconstruction; RS forcing chain (AbsoluteFloorClosure, Cost/FunctionalEquation, AlexanderDuality, etc.) has no opinion on attention mechanisms or monocular priors.","tokens_in":52143,"confidence":"high","tokens_out":441,"duration_ms":7525,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"EpiS reconstructs surfaces from sparse multi-view images by guiding fine-grained epipolar feature aggregation with coarse cost-volume features.","keywords":["neural surface reconstruction","epipolar geometry","sparse views","generalizable reconstruction","cost volume","SDF","monocular depth regularization","feature aggregation"],"falsifier":"On the DTU dataset using three input views, EpiS produces higher Chamfer distance or lower F-score than prior generalizable cost-volume baselines.","tokens_in":2589,"feed_emoji":"📐","tokens_out":690,"duration_ms":21182,"temperature":0.7,"pith_summary":"The paper tries to establish that explicitly incorporating epipolar geometry into a neural surface reconstruction pipeline overcomes the geometric ambiguity and information loss that plague cost-volume methods when inputs are limited to a few views. It replaces reliance on simple statistics like mean and variance with guided sampling of features along epipolar lines, fusion through an epipolar transformer, ray-wise aggregation into SDF-aware features, and scale-invariant regularization drawn from a pretrained monocular depth model. A sympathetic reader would care because sparse-view capture is the practical norm for many real scenes, yet existing generalizable approaches produce over-smoothed or incomplete surfaces; the new design promises accurate reconstruction without dense imagery or per-scene optimization. If correct, the approach would make high-fidelity surface modeling feasible from ordinary limited photo sets.","feed_headline":"Epipolar lines guide sparse-view surface reconstruction","feed_subtitle":"Framework aggregates features along epipolar lines to cut geometric ambiguity without dense inputs or per-scene tuning.","key_machinery":"Epipolar feature aggregation guided by cost-volume features, which samples and fuses view-dependent geometry along epipolar lines before producing SDF-aware outputs.","core_discovery":"The authors present EpiS as a generalizable framework that uses coarse cost-volume features to guide aggregation of fine-grained epipolar features sampled along corresponding epipolar lines across source views. An epipolar transformer fuses the multi-view information, followed by ray-wise aggregation to produce SDF-aware features for surface estimation. A geometry regularization strategy that leverages a pretrained monocular depth model through scale-invariant global and local constraints further mitigates information loss under sparse views.","pith_inferences":["The hybrid cost-volume plus epipolar strategy could transfer to other sparse multi-view tasks such as depth estimation or novel-view synthesis.","Similar monocular priors might regularize reconstruction in dynamic or non-rigid scenes where epipolar consistency still holds across frames.","The design implies that learned priors aligned with epipolar geometry can substitute for additional views in extremely sparse regimes."],"forward_implications":["Outperforms state-of-the-art generalizable surface reconstruction methods on DTU and BlendedMVS under sparse-view settings.","Maintains strong generalization without per-scene optimization.","Reduces over-smoothing by preserving view-dependent geometric structure that simple cost-volume statistics discard.","Handles occlusions and geometric ambiguity more effectively through explicit epipolar sampling and depth-based regularization."],"fun_headline_variants":["EpiS samples epipolar features for sparse surface reconstruction","Epipolar lines reduce ambiguity for neural surface reconstruction","Epipolar transformer fuses views for SDF-aware surface estimation","Geometry regularization aids EpiS in sparse multi-view settings","Cost volumes guide fine epipolar aggregation across source views"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Coarse cost-volume features can reliably guide fine-grained epipolar feature aggregation while a pretrained monocular depth model supplies unbiased scale-invariant constraints that align with multi-view epipolar geometry.","fun_headline_variants_meta":{"raw":{"variants":["EpiS samples epipolar features for sparse surface reconstruction","Epipolar lines reduce ambiguity for neural surface reconstruction","Epipolar transformer fuses views for SDF-aware surface estimation","Geometry regularization aids EpiS in sparse multi-view settings","Cost volumes guide fine epipolar aggregation across source views"]},"model":"grok-4.3","cost_usd":0.005176,"raw_usage":{"total_tokens":2503,"prompt_tokens":650,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":51762000,"prompt_tokens_details":{"text_tokens":650,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1778,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":650,"tokens_out":75,"duration_ms":12751,"temperature":1.0,"reasoning_tokens":1778,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-24T00:09:03.357112+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On the DTU dataset using three input views, EpiS produces higher Chamfer distance or lower F-score than prior generalizable cost-volume baselines.","supporting_citations":[],"review_version":1}