{"id":"616750d0-672c-4702-a67e-3f4cd71487c2","arxiv_id":"2506.06864","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A cGAN denoiser on three orthogonal planes plus an LDGCNN recognizer is reported to improve noisy 3D face recognition accuracy by up to 14.81% over prior methods on Bosphorus.","lead":"This paper combines a GAN-based point cloud denoiser that projects faces onto three orthogonal planes with a graph neural network recognizer, and reports accuracy gains up to 14.81% over prior methods on noisy Bosphorus point clouds. A generalist might read it because sensor noise is a practical obstacle for 3D face recognition, but the paper's evidence is thin: no error bars, no code, no ablation separating the recognizer from the denoiser.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No LDGCNN-without-denoising baseline is reported, so the 14.81% gain in Table 1 may come from the stronger backbone rather than the proposed cGAN-TOP denoiser.","rationale":"The paper's stated contribution is an end-to-end framework in which cGAN-TOP removes noise and LDGCNN recognizes faces, and the headline evidence is Table 1. The central inference is that the reported accuracy is due to the denoising model. That inference requires comparing the full pipeline against the recognition model applied to noisy data. The paper omits precisely that comparison. It compares against older recognition methods (LeD3D, FER) that use different backbones, so the improvement confounds the denoiser with the backbone. This is not an external-consensus issue; it is an internal control that must be present for the claimed attribution. The ablation table does not help because it varies only the discriminator-loss weights (λ1, λ2), not the presence of denoising. Therefore the 14.81% gain is not yet attributable to the proposed method. The reader identified this same gap among the weakest assumptions, and I focus on it as the single most load-bearing concern. The inverse-mapping issue is also significant, but it would not by itself eliminate the need for the backbone control. Since this missing control directly undermines the central quantitative claim, the rejection verdict stands.","tokens_in":6671,"tokens_out":3748,"duration_ms":39034,"concrete_test":"Re-run the recognition protocol of Section 3.1 with the same LDGCNN architecture and training settings on the noisy input point clouds directly, without cGAN-TOP denoising, for σ2 ∈ {4, 8, 16, 32, 64} under both Neutral and Random Settings. Compare the resulting accuracies to 'Ours' in Table 1. If LDGCNN-only accuracy at σ2 = 64 Random is at or near 51.61%, then the claimed gain cannot be attributed to the denoiser; if it is substantially lower, the missing baseline concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a maximum 14.81% improvement at σ2 = 64 under the Random Setting is established only by comparing 'Ours' (cGAN-TOP + LDGCNN) against LeD3D [16] and FER [15] in Table 1. Since LDGCNN is known to be a strong point-cloud classifier, it is possible that the entire accuracy advantage comes from the recognition backbone, not from the proposed denoising module. The paper reports no control experiment in which the same LDGCNN is trained and evaluated on noisy point clouds without cGAN-TOP. The ablation in Table 3 only varies the discriminator losses inside cGAN-TOP; it does not isolate the denoising stage. Without this baseline, the improvement cannot be attributed to the synergy of denoising and recognition, so the paper's primary quantitative claim is unsupported. A second internal gap—the unspecified inverse mapping f' from denoised 2D planes back to a sparse 3D point cloud in Section 2.2, including how multiple points projecting to the same pixel are handled—further undermines reproducibility, but even if that mapping were resolved, the missing backbone control would still leave the headline gain unexplained.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end 3D face recognition pipeline for noisy point clouds, consisting of a conditional GAN that denoises three orthogonal plane projections of the point cloud (cGAN-TOP) followed by an LDGCNN classifier. The method is evaluated on the Bosphorus dataset under five Gaussian noise levels and two train/test settings. The authors report accuracy improvements of up to 14.81% over prior methods and better Chamfer/Point-to-Mesh distances against dedicated denoising models.","tokens_in":6913,"tokens_out":5232,"duration_ms":58889,"significance":"If the claims are substantiated, the work would offer a practical way to leverage 2D GANs for point-cloud denoising in a recognition-driven loop, with a plausible dual-discriminator design (visual appearance plus recognition features). The paper explicitly targets an end-to-end integration, which is a worthwhile direction. However, the current evidence does not establish that the proposed denoiser is responsible for the reported gains, and the inverse mapping from denoised images back to point clouds is left unspecified. The manuscript also lacks reproducibility details such as error bars, training hyper-parameters, and precise noise injection protocols. These issues are load-bearing for the central claim.","major_comments":[{"comment":"The central quantitative claim of a maximum 14.81% accuracy gain at σ2=64 under the Random Setting is supported only by comparisons with LeD3D [16] and FER [15]. Since LDGCNN is a strong point-cloud recognition backbone, the reported advantage may be caused by the backbone rather than by the proposed cGAN-TOP denoiser. No control experiment is reported in which the same LDGCNN is trained and evaluated on noisy point clouds without the denoising stage. Please add this condition and report mean ± standard deviation over multiple runs with appropriate significance tests.","section":"Section 3.2, Table 1"},{"comment":"The reconstruction step, in which the three denoised 2D images are mapped back to a denoised 3D point cloud via the inverse function f′, is not specified. In particular, the paper does not explain how points that project to the same pixel in a given plane are resolved, how the three independently denoised coordinate images are fused into a consistent point set, or what happens when the denoised images produce different point counts in the three planes. This missing definition makes the method irreproducible and raises the possibility of systematic loss of geometric information. Please provide a precise algorithm for f′ and discuss its behavior with respect to occlusions, multiplicity, and information preservation.","section":"Section 2.2, Fig. 2"},{"comment":"The ablation study varies which discriminator losses are used, but it never removes the cGAN-TOP denoiser entirely; feeding noisy point clouds directly to LDGCNN under identical training conditions is not reported. Consequently, the experiment does not isolate the contribution of the denoising module. In addition, the hyper-parameters λ1=0.67 and λ2=0.33 are said to be tuned to achieve optimal performance on the same dataset, but no validation split or sensitivity analysis is described. Please clarify how these values were selected and report accuracy over a range of λ1/λ2 values to rule out overfitting to the test set.","section":"Section 2.2 and Table 3"},{"comment":"The experimental protocol reports only single-run accuracy numbers without standard deviations or the number of repeated trials. For several noise levels the differences between methods are small (e.g., Neutral Setting at σ2=4: Ours 98.74 vs FER 96.88), so it is impossible to assess whether the observed improvements are statistically reliable. Please provide mean ± std over at least three independent runs, or explicitly state whether training is deterministic and the test set is fixed so that repeated runs are unnecessary. The paper also omits details on the number of training/test subjects, the exact noise injection procedure (e.g., per-coordinate independent Gaussian noise or along surface normals), and the point-cloud alignment protocol.","section":"Section 3.1"}],"minor_comments":[{"comment":"The abstract reports a maximum gain of 14.81% without stating that it occurs at σ2=64 under the Random Setting; please make this condition explicit.","section":"Abstract"},{"comment":"The notation f: (x,y) → (x′, y′) is introduced but never used or explained; please clarify what coordinate transformation this represents and how it relates to the pixel grid.","section":"Section 2.2"},{"comment":"The phrase 'the three denoised coordinates are then combined' is ambiguous in light of the missing inverse mapping; consider rewriting once the reconstruction procedure is specified.","section":"Section 2.1"},{"comment":"Please specify how Chamfer distance and Point-to-Mesh distance are computed, including point-density normalization and whether the denoised point clouds are subsampled or resized to a common resolution.","section":"Table 2"},{"comment":"The text says the proposed method 'shows advantages at a high noise level and comparable performance to SB at a low noise level,' but Table 2 does not include confidence intervals or significance tests for the CD/P2M metrics; please address this.","section":"Section 3.2"},{"comment":"There are minor typographical and consistency issues: 'VFD' and 'LRFD' in Table 3 do not match the 'VAD' and 'RFD' abbreviations used elsewhere; 'visually' is misspelled as 'visual' in Section 2.2; and 'plan' appears instead of 'plane' in the conclusion.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like a conference paper in both format and depth. The two load-bearing gaps—the absence of an LDGCNN-without-denoising baseline and the unspecified inverse mapping—are fixable in principle with additional experiments and a detailed algorithmic description. If the authors can supply those, the paper might become publishable; in its current form the central claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The cGAN-TOP idea is the genuinely new part: project a noisy point cloud onto three orthogonal planes, treat the out-of-plane coordinate as a gray value, denoise each plane with a conditional GAN, then reconstruct. That is a reasonable engineering move for thin, surface-like face clouds, and the two-discriminator design (visual appearance plus recognition feature) is a sensible way to tie denoising to downstream recognition. The paper also reports consistent gains over LeD3D and FER across noise levels, and the Chamfer/P2M numbers in Table 2 are at least competitive with dedicated denoisers at high noise.\n\nThe soft spot is that the central quantitative claim is not actually supported. Table 1 compares \"Ours\" (cGAN-TOP + LDGCNN) against LeD3D and FER, but never runs LDGCNN alone on the noisy clouds. Since LDGCNN is a strong point-cloud classifier, the 14.81% gain could easily come from the backbone rather than the proposed denoiser. The ablation in Table 3 only varies the discriminator losses inside cGAN-TOP; it does not remove the denoiser entirely. That missing control is the load-bearing flaw, and it is fixable.\n\nThere are secondary gaps. The inverse mapping f' from the denoised 2D planes back to a sparse 3D cloud is left unspecified, including what happens when multiple points project to the same pixel. The paper reports single-run accuracies with no error bars or significance tests. The loss weights λ1=0.67 and λ2=0.33 are tuned on the same test data, which is mild but worth stating. No code is provided. None of these are fatal on their own, but they add up.\n\nOn the positive side, the self-citations are background; there is no circularity load. The paper is clearly written and the method is reproducible in principle, though not in practice without the missing details.\n\nMy take: the method has potential, but as submitted this version does not support the headline claim. I would not cite the numbers yet. If I were handling it, I would send it back for revision rather than desk reject, because the missing baseline and the inverse-mapping specification are both directly addressable and the core idea is not obviously wrong. A serious referee could quickly tell whether the gains survive once LDGCNN is run without denoising. So: yes to peer review, but with a clear request for that control experiment.","headline":"The three-plane cGAN denoising idea is sensible, but the headline 14.81% gain is not supported because no LDGCNN-without-denoising baseline is reported.","tokens_in":7421,"tokens_out":2453,"would_cite":false,"duration_ms":29209,"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":"One denoiser lifts noisy 3D face recognition by 14.81%.","keywords":["3D face recognition","point cloud denoising","conditional GAN","three orthogonal planes","LDGCNN","Bosphorus dataset","noisy point cloud","end-to-end learning"],"falsifier":"Train the same LDGCNN recognizer on the noisy point clouds without any denoising, using the identical training protocol; if its accuracy matches or exceeds the cGAN-TOP pipeline at σ2=64 in the Random Setting, the denoiser is not the cause of the 14.81% gain, and also inspect the reconstruction step to see whether multiple points projected into the same pixel are handled without collapse or duplication.","tokens_in":6433,"feed_emoji":"👤","tokens_out":4324,"duration_ms":33711,"temperature":0.7,"pith_summary":"This paper proposes an end-to-end system that first denoises a noisy 3D face point cloud and then recognizes the identity from the cleaned cloud. The denoiser, cGAN-TOP, projects the cloud onto three orthogonal planes, treats the coordinate perpendicular to each plane as a gray value, and runs a conditional GAN per plane to remove noise. A Recognition Feature Discriminator pushes the denoised cloud to retain identity-discriminative features, while a Visual Appearance Discriminator keeps it geometrically clean. On the Bosphorus dataset the method beats the previous baselines at every tested noise level, with the largest gain of 14.81% accuracy at the highest noise level under the Random Setting.","feed_headline":"One denoiser lifts noisy 3D face recognition by 14.81%","feed_subtitle":"Cleaning point clouds on three planes before recognition beats prior models at every noise level.","key_machinery":"cGAN-TOP (Conditional GAN on Three Orthogonal Planes): the point cloud is projected onto xy, yz, and zx planes; in each plane the orthogonal coordinate becomes the gray value and a UNet-style generator denoises that 2D image. Two discriminators guide training: a Visual Appearance Discriminator (DCGAN-like) that enforces geometric cleanness and a Recognition Feature Discriminator that enforces identity-relevant features. The denoised coordinates are mapped back and combined into a point cloud, which is then classified by an adapted LDGCNN that links hierarchical local and multi-scale neighboring features.","core_discovery":"The central claim is that denoising and recognition should be trained together rather than treated as separate stages, and that denoising can be done effectively on three orthogonal 2D projections of a point cloud by treating each spatial coordinate as a pixel gray value. In this design, cGAN-TOP removes noise in each coordinate channel, and the two discriminators together guarantee both visual cleanness and recognition suitability of the output. An adapted LDGCNN then classifies the reconstructed point cloud. The paper reports that this integrated pipeline outperforms LeD3D and FER under all five Gaussian noise levels on Bosphorus, with accuracy gains up to 14.81% at σ2=64 in the Random Setting, and also produces lower Chamfer and Point-to-Mesh distances than PCNet, DMR, and SB at high noise levels.","pith_inferences":["The paper does not report recognition with LDGCNN alone on noisy clouds; an obvious test is whether the 14.81% gain survives when the denoiser is removed and the same backbone is used, which would isolate the denoiser's contribution.","The inverse mapping from denoised 2D images back to a 3D point cloud is not specified for pixels receiving multiple points or no points; resolving this ambiguity could affect the geometry-preservation claim.","If the same cGAN-TOP design works for other sparse 3D recognition problems, such as object class recognition on LiDAR-style clouds, the orthogonal-projection denoising idea becomes a general preprocessing layer.","The reported gains are on one dataset with injected Gaussian noise; testing on real sensor noise from depth-camera scans would show whether the advantage transfers."],"forward_implications":["If the cGAN-TOP denoiser is the source of the gain, then joint denoising-recognition training on orthogonal projections is a viable general recipe for other sparse 3D recognition tasks on noisy sensor data.","The Recognition Feature Discriminator is what ties denoising to recognition; the ablation suggests dropping either discriminator costs accuracy, so recognition-aware denoising is necessary for the reported results.","At high noise levels the method improves both recognition accuracy and geometric cleanness, indicating the two objectives are not in conflict in this framework.","The paper claims the approach suits thin, surface-like point clouds where 3D graph convolutions struggle, such as facial scans from consumer depth sensors."],"supporting_citations":[{"why":"Provides the LDGCNN backbone adapted for face recognition, which carries the recognition module.","marker":"[6]"},{"why":"FER is the state-of-the-art baseline it must beat on noisy point clouds.","marker":"[15]"},{"why":"LeD3D is the lightweight low-quality 3D face recognition baseline and the noise injection protocol follows its setting.","marker":"[16]"},{"why":"Source of the conditional GAN formulation used for denoising in each of the three planes.","marker":"[18]"},{"why":"Supplies the Bosphorus dataset used for all experiments.","marker":"[19]"},{"why":"UNet architecture used by each generator in cGAN-TOP.","marker":"[26]"},{"why":"DCGAN design used for the Visual Appearance Discriminator.","marker":"[27]"}],"fun_headline_variants":["Joint denoising+recognition lifts 3D face accuracy by 14.81%","cGAN-TOP denoiser boosts noisy 3D face recognition by 14.81%","Cleaning point clouds on three planes beats prior models","Integrated denoise and recognize improves 3D faces by 14.81%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported improvement assumes that mapping the denoised 2D planes back into a sparse 3D point cloud via the unspecified inverse function preserves enough geometric structure for the recognizer, and that the measured gain comes from the cGAN-TOP denoiser rather than from the stronger LDGCNN backbone alone.","fun_headline_variants_meta":{"raw":{"variants":["Joint denoising+recognition lifts 3D face accuracy by 14.81%","cGAN-TOP denoiser boosts noisy 3D face recognition by 14.81%","Cleaning point clouds on three planes beats prior models","Integrated denoise and recognize improves 3D faces by 14.81%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000136,"raw_usage":{"total_tokens":1103,"prompt_tokens":861,"completion_tokens":242,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":154}},"tokens_in":477,"tokens_out":242,"duration_ms":3428,"temperature":1.0,"reasoning_tokens":154,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:46:40.135331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same LDGCNN recognizer on the noisy point clouds without any denoising, using the identical training protocol; if its accuracy matches or exceeds the cGAN-TOP pipeline at σ2=64 in the Random Setting, the denoiser is not the cause of the 14.81% gain, and also inspect the reconstruction step to see whether multiple points projected into the same pixel are handled without collapse or duplication.","supporting_citations":[{"cited_title":"Learning from millions of 3D scans for large-scale 3D face recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the LDGCNN backbone adapted for face recognition, which carries the recognition module."},{"cited_title":"VERAM: view-enhanced recurrent attention model for 3D shape classifi- cation,","cited_arxiv_id":null,"evidence_quote":"FER is the state-of-the-art baseline it must beat on noisy point clouds."},{"cited_title":"Learning directly from synthetic point clouds for","cited_arxiv_id":null,"evidence_quote":"Source of the conditional GAN formulation used for denoising in each of the three planes."},{"cited_title":"Repcd- net: Feature-aware recurrent point cloud denoising network,","cited_arxiv_id":null,"evidence_quote":"UNet architecture used by each generator in cGAN-TOP."},{"cited_title":"Three-dimensional point cloud denoising via a gravitational feature function,","cited_arxiv_id":null,"evidence_quote":"DCGAN design used for the Visual Appearance Discriminator."}],"review_version":1}