{"id":"fdcc6ee3-303e-4e70-a431-66380f4fab9f","arxiv_id":"2504.18910","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A forest of two-node graphs over corresponding parent-child face components, combined with a gradually introduced center loss, reports state-of-the-art accuracy on KinFaceW-II.","lead":"This paper pairs up face parts from a parent and a child, connects matching parts in small graphs, and trains a graph network to decide whether they are family. It reports the best average accuracy so far on KinFaceW-II, but that result is weakened by test-set hyperparameter selection.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CNN embeddings are trained on all KinFaceW identities, including FNN test-fold identities, so the reported SOTA may reflect identity memorization rather than kinship generalization.","rationale":"The reader's weakest assumption is precisely that identities and images used to train the CNN embedding extractors do not leak into the FNN test folds. The paper's own Section 4.2 describes a per-identity random split of augmented images into 14 training and 3 testing images, which includes all identities—so every FNN test identity is represented in CNN training. This is a direct violation of the subject-disjoint evaluation needed for a valid kinship generalization claim. The concern is load-bearing because the central claim is an empirical SOTA result; if the CNN has memorized test identities, the reported numbers are not meaningful for the general kinship verification task. The proposed concrete test—an identity-disjoint retraining of the CNNs—directly checks whether the reported accuracy is an artifact of identity leakage. The reader's rationale also correctly identifies secondary issues (test-based hyperparameter selection in Tables 6 and 8, zero-valued loss equations 13–14, missing error bars), but the identity leakage is the most fundamental and would independently invalidate the claim if confirmed. I find no reason to change the reader's REJECT verdict; the concern is serious and concrete, and the authors have not provided the necessary protocol details or experiments to rule it out.","tokens_in":18174,"tokens_out":3758,"duration_ms":40960,"concrete_test":"Re-run the full pipeline under a subject-disjoint protocol. For each of the five FNN folds, train every CNN embedding network (ResNet-18, ERN, VGG as used in the winning configuration) using only augmented images from identities whose kin pairs are in the FNN training folds; use images from the held-out identities only for the FNN test fold. Exclude all images of test-fold identities from CNN training and from any GFP-GAN fine-tuning. Then compute the FNN mean accuracy for KinFaceW-II and compare to the reported 93.8. If accuracy drops by more than 2–3 points (or a statistical test shows a significant difference), the leakage explanation is confirmed; if accuracy is unchanged within fold-to-fold variance, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the reported KinFaceW-II accuracy (93.8 mean, Table 4) is a valid out-of-sample result. That validity depends on the CNN embedding networks not having been trained on images of the identities that appear in the FNN test folds. Section 4.2 states: 'All images of every identity are randomly split into 14 and 3 for training and testing, respectively' for the CNN networks. Because every identity is included in this split, every FNN test identity has augmented images (including masked and GFP-GAN restored versions) in CNN training. The paper never states that a subject-disjoint split was used for CNN pretraining, and the described procedure is the opposite. The FNN then computes embeddings from these CNNs for the same identities it later classifies. If the CNN embeddings encode identity-specific features, the kinship classifier can exploit correlations between identity features and kinship, inflating accuracy beyond true generalization. This is particularly dangerous here because the CNN is trained on heavily augmented and synthetic images of the exact test identities. Without an identity-disjoint CNN training protocol, the SOTA claim is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Forest Neural Network (FNN) for kinship verification. It trains several CNN embedding extractors (ResNet-18, an attention-based Extended Residual Network, and VGG-Face) on augmented, masked, and GFP-GAN-restored KinFaceW images, extracts embeddings from nine facial patches, arranges them into paired two-node graphs forming a forest, and processes the forest with Residual Gated Graph Convolutional layers. The model then concatenates per-layer mean readouts for parent and child, applies four feature combinations, and classifies with a three-layer MLP trained with a fusion of classification, family-ID, triplet, cross-generation, direction, and temperature-scheduled center losses. On KinFaceW-II the method reports a mean accuracy of 93.8%, the best in the comparison table, with a 1.6-point gain over the next best method; on KinFaceW-I it reports 82.09%, near the best.","tokens_in":18299,"tokens_out":10669,"duration_ms":101967,"significance":"If the empirical results are valid, the paper makes a useful contribution by showing that face-representation-based kinship verification can compete with joint-representation methods when a forest of paired graphs exchanges information across facial components and when multiple pretrained embeddings are combined. Strengths include the public code release, a wide comparison table, model-size comparison, and ablations for CNN combinations and the center loss. The main value is the architectural idea of applying GNN message passing to paired facial components. However, the reported SOTA numbers are not supported by the current evaluation protocol because of test-set selection of hyperparameters, a potential identity-level leak in CNN pretraining, missing variance estimates, and an inconsistency between the 'good configurations' tables and the headline results.","major_comments":[{"comment":"Section 4.2 states that 'All images of every identity are randomly split into 14 and 3 for training and testing, respectively' for the CNN networks, meaning that every identity that later appears in an FNN test fold has augmented, masked, and GFP-GAN-restored images in the CNN training set. The FNN evaluation in Section 4.3 then uses embeddings of these exact identities. The manuscript nowhere states the CNN training loss nor that an identity-disjoint split was used. If the CNN is trained with identity supervision, the downstream classifier can exploit identity memorization rather than kinship generalization, which would inflate the headline accuracy. Please specify the CNN training loss and re-run the evaluation with an identity-disjoint protocol for the embedding networks.","section":"Sections 4.2 and 4.3"},{"comment":"The hyperparameters α, H1, and H2 are tuned by the reported accuracies in Tables 6 and 8, which are computed on the same five-fold test partitions used to produce Table 4. The final 'Ensemble of good configurations' is therefore selected on the test folds, so the headline 93.8% is a fitted maximum over the searched configurations rather than an independent estimate. Moreover, the final mean accuracies (82.09 and 93.8) exceed every accuracy listed in the corresponding 'good configurations' tables (maxima 80.93 and 93.3), which is inconsistent unless the ensemble procedure is different from the configurations listed; please clarify the ensemble definition and use nested cross-validation or a held-out validation set for model selection.","section":"Section 4.5.3, Tables 6, 8, and 4"},{"comment":"Equations (13) and (14) define Lpos = (1/n) Σ y_i ||F^p_i - F^p_i||_2^2 and Lneg = (1/n) Σ (1-y_i) ||F^p_i - F^p_i||_2^2, both of which are identically zero because each norm is taken between F^p_i and itself. If implemented as written, the cross-generation divergence loss in Eq. (12) contributes nothing, contrary to the description in Section 3.7.3 and to the MSE terms in Algorithm 1. The intended expressions presumably use F^c_i; please correct the equations and verify that the reported results used the corrected loss.","section":"Equations (13) and (14)"},{"comment":"All accuracy numbers are point estimates without standard deviations or per-fold breakdowns. Because the KinFaceW-I test folds contain only a few dozen pairs per relation (e.g., about 31 father-son pairs per fold), a difference of 1.6 points may be within sampling variability. Please report per-fold accuracies, means with standard deviations, and ideally a significance test before claiming a state-of-the-art margin.","section":"Tables 4, 5, and 7"}],"minor_comments":[{"comment":"The text reports an improvement of '3.07' and then repeats 'in the KinFaceW-I dataset' for the second set of improvements (1.4, 2.4, 1.4, 0.6); the second set appears to refer to KinFaceW-II, not KinFaceW-I.","section":"Section 4.5.2"},{"comment":"The configuration (1.05, 256, 16) is listed twice with different accuracies (93.15 and 92.95); one of these entries is presumably a typo.","section":"Table 8"},{"comment":"The summation is written as '9X 1'; it should be the sum over i=1 to 9 of f^l_pi.","section":"Equation (5)"},{"comment":"There are several typos: 'ReseNet-18' in Section 3.1, 'Deatails' in the Section 3.5 heading, 'Dawn Scale' and 'Hue Santuration' in Table 3, and 'FFN' in Figure 1 where 'FNN' is intended.","section":"Throughout"},{"comment":"The phrase 'in companion with' should be 'in combination with'.","section":"Section 4.5.4"},{"comment":"No random seed is reported, so the experiments are not exactly reproducible; please report seeds or averaged runs over multiple seeds.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper reports state-of-the-art kinship verification on KinFaceW-II (93.8 mean) and near-best on KinFaceW-I, but the evaluation does not support the claim. The architecture is a reasonable assembly of known parts—multiple CNN embeddings (ResNet, ERN, VGG), paired two-node graphs with residual gated updates, and a gradually engaged center loss—and the writing is clear. The code is promised; the model-size table is useful. What's genuinely new is the specific combination, not a new principle.\n\nThe problems are in the protocol. Section 4.2 says all images of every identity are randomly split into 14 for training and 3 for testing the CNN networks. That means every identity appearing in the FNN test folds also had augmented, masked, and GFP-GAN-restored images in CNN training. The FNN then uses embeddings from these CNNs to classify kin/non-kin pairs for those same identities. The paper never states that an identity-disjoint split was used, and the described procedure is the opposite. With face-embedding networks, identity memorization can easily inflate kinship accuracy beyond true generalization. This is load-bearing.\n\nSecond, hyperparameters (alpha, H1, H2) and the ensemble configuration were chosen based on test-fold accuracy (Tables 6 and 8). The headline number is a selected maximum, not an out-of-sample result. Third, Equations 13 and 14 define Lpos and Lneg as zero as written (they subtract F_p from F_p); a typo, but it makes the loss description unusable. Fourth, no error bars or significance tests are reported.\n\nThe 'first GNN' claim is also overstated; H-RGN already does graph-based reasoning, even if not exactly a GNN. None of these issues is fatal to the idea, but together they make the empirical claim unsupported.\n\nThis paper is for readers in face-based kinship verification, and for anyone who wants a case study in how evaluation protocol can break a benchmark claim. I would not cite it as a SOTA result. It does deserve peer review, though—the method is plausible and the flaws are fixable. A resubmission with identity-disjoint CNN pretraining, validation-based hyperparameter selection, corrected loss equations, and error bars would be a different paper.","headline":"Plausible architecture, unsupported SOTA: claimed KinFaceW-II gains likely inflated by identity leak and test-set hyperparameter selection.","tokens_in":18959,"tokens_out":3998,"would_cite":false,"duration_ms":36078,"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":"A forest of paired two-node graphs between corresponding facial components lets face-representation methods match or beat joint-representation kinship verification.","keywords":["Kinship Verification","Forest Neural Network","Graph Neural Network","Face Representation","Center Loss","KinFaceW-I","KinFaceW-II","Residual Gated Graph ConvNets"],"falsifier":"Re-run the full pipeline under a strict identity-disjoint protocol: train every CNN embedding extractor only on identities that do not appear in any of the five FNN test folds, then compare the mean KinFaceW-II accuracy. If the accuracy drops materially below the reported 93.8%, or no longer beats D4ML, the central claim of competitive face-representation kinship verification would be unsupported.","tokens_in":17864,"feed_emoji":"🧬","tokens_out":8123,"duration_ms":79200,"temperature":0.7,"pith_summary":"Kinship verification usually asks whether two face images are parent and child, and the strongest recent systems learn a joint representation of the two faces from scratch. This paper tries to show that the older strategy of using separate face representations can be just as good if the representations are organized as a forest of paired component graphs. Nine facial regions (whole face, left eye, right eye, nose, mouth, and four masked faces) are embedded by several pretrained convolutional networks, and each corresponding parent-child pair becomes a two-node graph whose nodes exchange gated messages. On KinFaceW-II the method reports 93.8% mean accuracy, the best among the compared methods and about 1.6 points above the previous best; on KinFaceW-I it reaches 82.09%, close to the best, with the top mother-daughter score. If correct, the result means face-representation pipelines can compete with joint-representation ones while using far fewer parameters.","feed_headline":"Forest of paired face graphs tops KinFaceW-II","feed_subtitle":"Separate face-region embeddings reach 93.8 percent accuracy by pairing components into a graph forest.","key_machinery":"The central object is the Forest Neural Network (FNN): a forest of nine disjoint two-node graphs, each joining a parent's facial-component embedding to the child's corresponding embedding with a bidirectional edge. The message-passing layer is the Residual Gated Graph ConvNet, simplified to two nodes: the parent update is $\\mathbf{h}_p^{l+1} = \\mathbf{h}_p^l + f(\\mathbf{A}\\mathbf{h}_p^l + \\eta(e)\\odot \\mathbf{B}\\mathbf{h}_c^l)$ with gate $\\eta(e_p)=\\sigma(e_p)/(\\sigma(e_p)+\\sigma(e_c))$, so each node is updated by combining its own state with its partner's state, scaled by a learnable gate. After each layer the nine parent representations are averaged (readout), and the averages from four layers are concatenated into final parent/child features; the classifier then concatenates four feature combinations and uses a loss mixture whose center-loss term is introduced gradually via a temperature schedule $\\alpha^t$. This machinery is what lets separate face representations carry kinship information: correspondence is imposed by graph structure rather than by joint training from scratch.","core_discovery":"The paper's central claim is that kinship can be verified from independent face embeddings without joint parent-child training, provided the component structure of faces is respected. It constructs nine paired graphs, each linking one facial component of the parent to the same component of the child, and processes the forest with a residual gated graph convnet in which the parent and child nodes exchange information through gated edges. A mean readout across the nine graphs at each layer, concatenated over layers, forms parent and child feature vectors; these are combined by sum, product, squared difference, and difference-of-squares features and passed to a small classifier. With a five-term loss that adds a gradually weighted center loss, the method reaches 93.8% mean accuracy on KinFaceW-II, the highest in the comparison table, and 82.09% on KinFaceW-I, near the best and the highest for mother-daughter. The authors state this as evidence that face-representation approaches can match joint-representation state-of-the-art.","pith_inferences":["If the evaluation is later confirmed to be identity-disjoint, the component-paired forest predicts that kinship signal is not uniform across the face: ablating individual graphs should change accuracy more than removing others, a testable extension the paper does not report.","The same two-node graph construction could be applied to other pairwise verification tasks where two images share corresponding parts, such as age-progressed face verification or sibling verification, with the gradual center-loss schedule acting as a regularizer.","Because the gain is larger on KinFaceW-II (250 pairs per relationship) than on KinFaceW-I (about 130 pairs), the method may benefit more as kinship training data grows; evaluating on a larger benchmark would reveal whether the advantage scales.","If the reported accuracy depends on CNN embeddings trained on the same identities used in FNN test folds, an unstated protocol detail, the method's true generalization may be lower; the released code makes this directly checkable."],"forward_implications":["Face-representation kinship systems can reach state-of-the-art accuracy if corresponding facial components are paired explicitly, so joint representation from scratch is not the only route.","Multiple pretrained embeddings at different resolutions act as complementary experts; combining ResNet, ERN, and VGG embeddings improves accuracy over any single embedding in the reported tables.","The center loss only helps when applied to a hidden layer of the kinship classifier and when its influence grows during training; with the temperature schedule it contributes roughly 1.45 points on KinFaceW-II and about 3.07 on KinFaceW-I.","The proposed FNN classifier is substantially smaller than the joint-representation D4ML model (about 19M versus 91M parameters on KinFaceW-I), suggesting a practical size advantage."],"supporting_citations":[{"why":"Supplies the KinFaceW-I/II datasets and the five-fold kin/non-kin evaluation protocol used for all comparisons.","marker":"[6]"},{"why":"The D4ML baseline that this method compares against and the source of the feature-combination and multi-loss design.","marker":"[9]"},{"why":"The graph-oriented kinship baseline whose star-shaped-to-hierarchical reasoning the FNN extends with genuine GNN message passing.","marker":"[8]"},{"why":"Provides the Residual Gated Graph ConvNet operator used as the two-node message-passing layer in every paired graph.","marker":"[48]"},{"why":"GFP-GAN blind face restoration used to create high-resolution 256x256 images that supply additional embedding networks.","marker":"[44]"},{"why":"Introduces the center loss that the paper integrates into kinship classification with a gradual temperature schedule.","marker":"[50]"},{"why":"Motivates masked face components and part-aware attention, used to form the nine patches and the attention blocks.","marker":"[35]"},{"why":"SPIGA landmark detector used to locate facial components for building the paired graphs.","marker":"[47]"}],"fun_headline_variants":["Graph forest tops KinFaceW-II without joint training","Forest net of paired face graphs hits 93.8% kin","Separate face embeddings do kin via graph forest","No joint model: forest beats KinFaceW-II baseline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the identities used to train the CNN embedding extractors were never part of the kinship-classifier test folds; the paper does not describe such an identity-disjoint split, and without it the reported accuracy could reflect identity memorization rather than kinship generalization.","fun_headline_variants_meta":{"raw":{"variants":["Graph forest tops KinFaceW-II without joint training","Forest net of paired face graphs hits 93.8% kin","Separate face embeddings do kin via graph forest","No joint model: forest beats KinFaceW-II baseline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000639,"raw_usage":{"total_tokens":2903,"prompt_tokens":865,"completion_tokens":2038,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":1984}},"tokens_in":481,"tokens_out":2038,"duration_ms":14528,"temperature":1.0,"reasoning_tokens":1984,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:07:04.380974+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full pipeline under a strict identity-disjoint protocol: train every CNN embedding extractor only on identities that do not appear in any of the five FNN test folds, then compare the mean KinFaceW-II accuracy. If the accuracy drops materially below the reported 93.8%, or no longer beats D4ML, the central claim of competitive face-representation kinship verification would be unsupported.","supporting_citations":[{"cited_title":"IEEE transactions on pattern analysis and machine intelligence 36(2), 331–345 (2013)","cited_arxiv_id":null,"evidence_quote":"Supplies the KinFaceW-I/II datasets and the five-fold kin/non-kin evaluation protocol used for all comparisons."},{"cited_title":"ACM Transactions on Multimedia Computing, Communications and Applications 19(1s), 1–19 (2023)","cited_arxiv_id":null,"evidence_quote":"The D4ML baseline that this method compares against and the source of the feature-combination and multi-loss design."},{"cited_title":"IEEE Transactions on image processing 30, 4947–4961 (2021)","cited_arxiv_id":null,"evidence_quote":"The graph-oriented kinship baseline whose star-shaped-to-hierarchical reasoning the FNN extends with genuine GNN message passing."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"GFP-GAN blind face restoration used to create high-resolution 256x256 images that supply additional embedding networks."},{"cited_title":"In: Computer vision–ECCV 2016: 14th European Con- ference, Amsterdam, the Netherlands, October 11–14, 2016, Proceedings, Part VII 14, pp","cited_arxiv_id":null,"evidence_quote":"Introduces the center loss that the paper integrates into kinship classification with a gradual temperature schedule."},{"cited_title":"Pattern Recognition Letters 128, 169–175 (2019)","cited_arxiv_id":null,"evidence_quote":"Motivates masked face components and part-aware attention, used to form the nine patches and the attention blocks."},{"cited_title":"In: 33rd British Machine Vision Con- ference 2022, BMVC 2022, London, UK, November 21-24, 2022, p","cited_arxiv_id":null,"evidence_quote":"SPIGA landmark detector used to locate facial components for building the paired graphs."}],"review_version":1}