{"id":"9e1c1f55-9933-48e0-a1be-24e5d9673488","arxiv_id":"2505.09073","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A 2D-3D domain adaptation framework with shared attention and a joint entropy regularizer improves profile-view face recognition, though the entropy loss is mis-specified.","lead":"This paper combines shared attention and a joint entropy loss between 2D facial images and 3D point clouds during training, aiming to make 2D face recognition more accurate at extreme pose differences. It reports gains on FaceScape and ARL-VTF, but the joint entropy loss as written does not actually couple the two modalities.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The joint entropy loss in Eq. 6 factorizes into marginal entropies, so it cannot enforce 2D-3D attention consistency; the paper's central mechanism is unsupported.","rationale":"I read the paper in good faith. The strongest claim is that JAM and JE jointly yield pose-robust 2D recognition by emphasizing correlated 2D-3D patterns and promoting consistency between the two attention maps. For that to hold, the JE term must couple the two attention maps. It does not: the all-pairs histogram in Eq. 6 factorizes, so the joint entropy is exactly the sum of independent marginal entropies. This is a mathematical property and does not depend on the empirical tables. The reader's weakest_assumption identifies precisely this factorization, and I agree with it. The empirical results are not dismissed, but they are not accompanied by code or binning details, and the one ablation comparing JAM+JE against JAM alone shows a mixed effect: on FaceScape the 90+ column improves by 0.41% while average TAR decreases by 0.73%; on ARL-VTF the pose row improves, which is consistent with a general low-entropy regularizer rather than a cross-modal coupling. Without a valid coupling loss, the stated mechanism for the central contribution is unsupported, so the prevailing REJECT verdict stands unchanged.","tokens_in":17730,"tokens_out":3407,"duration_ms":32224,"concrete_test":"Obtain the training code or a precise histogram-construction description and evaluate LJE on a fixed checkpoint using (a) the all-pairs definition of Eq. 6 and (b) a corresponding-element definition, P(a,b) = (1/N) sum_i I{A2D(i)=a, A3D(i)=b}. If the two losses are equal, the implementation is not enforcing any cross-modal dependence. Additionally, re-run the ARL-VTF pose row of Table III with LJE replaced by H(A2D)+H(A3D); if the reported profile gain persists unchanged, the improvement is not caused by joint consistency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LJE promotes consistency between A2D and A3D. But the distribution defined in Eq. 6 is P(a2D,a3D) = (1/N^2) sum_{i,j} I{A2D(i)=a2D} I{A3D(j)=a3D} = [count2D(a2D)/N] [count3D(a3D)/N], a product of marginal histograms. Therefore H(A2D,A3D) = H(A2D) + H(A3D). There is no coupling between corresponding attention locations: the sum runs over all i,j pairs, so spatial position is discarded. Minimizing LJE only makes each attention map more peaked by lowering its marginal entropy. It does not make the 2D and 3D maps agree with one another, and it does not 'promote consistency among intersecting 2D and 3D representations' as claimed in Sections I, III, and IV-E. Because this is the stated mechanism behind the main pose-invariance contribution, the abstract's causal story is not supported by the mathematics. The empirical gains in Table I and Table III could still reflect the JAM shared-parameter attention plus a low-entropy regularizer, but they cannot be attributed to cross-modal consistency via Eq. 6. No code or binning details are provided, so it is also possible the implementation silently used a different paired-index histogram; that would need to be disclosed. As written, the JE contribution is internally inconsistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a domain-adaptive framework for pose-robust 2D face recognition. It combines a shared 2D-3D Joint Attention Mapping (JAM), which uses shared query/key convolutions to compute attention for both 2D images and 3D point clouds, with a Joint Entropy (JE) regularizing loss that is claimed to promote consistency between the 2D and 3D attention maps. The framework is trained on FaceScape, which provides paired 2D and 3D data, and evaluated on FaceScape and the 2D-only ARL-VTF dataset. The authors report TAR@1% FAR improvements of at least 7.1% and 1.57% for extreme profile poses on FaceScape and ARL-VTF, respectively, and present ablations showing that JAM and JE contribute to the results.","tokens_in":18058,"tokens_out":3922,"duration_ms":39309,"significance":"If the stated mechanism were correct, the paper would offer a practically useful way to inject 3D pose-invariance into 2D face recognition without requiring 3D data at inference. The cross-dataset evaluation on ARL-VTF is a genuine strength, and the JAM idea of computing attention with shared 2D/3D parameters is interesting. However, the central theoretical claim is not supported by the paper's own mathematics: the joint entropy loss defined in Eq. (6) factorizes into the sum of two marginal entropies and therefore cannot enforce any correspondence between 2D and 3D attention locations. This invalidates the abstract's causal story and the motivation in Sections III and IV. The empirical gains in Tables I and III could still stem from JAM plus a low-entropy regularizer, but the paper as written does not establish the claimed cross-modal consistency mechanism. No code or binning details are provided, making the loss irreproducible as defined.","major_comments":[{"comment":"The joint distribution P(a2D,a3D) is defined as a normalized sum over all pairs of positions (i,j), i.e., P(a2D,a3D) = (1/N^2) * sum_{i,j} I{A2D(i)=a2D} I{A3D(j)=a3D}. This factorizes exactly as P(a2D) * P(a3D), where P(a2D) = (1/N) sum_i I{A2D(i)=a2D} and similarly for P(a3D). Therefore H(A2D,A3D) = H(A2D) + H(A3D), and minimizing LJE is equivalent to independently minimizing the marginal entropies of the two attention maps. It does not penalize differences or inconsistencies between corresponding 2D and 3D attention locations. The claims in the Abstract, Section III, and Section IV-E that JE 'penalizes differences or inconsistencies between the feature maps' and 'promotes consistency among intersecting 2D and 3D representations' are thus not supported by the equation as written. This is a load-bearing issue because the paper attributes its pose-invariance gains to this loss.","section":"Section IV-E, Eq. (6)"},{"comment":"All quantitative results are reported as point estimates averaged over 'three independent training cycles,' but no standard deviations, per-fold values, or significance tests are provided. The headline ARL-VTF pose improvement over CrossPoint is 1.571 percentage points (54.678 vs. 53.107), which is small enough that run-to-run variation could change the conclusion. The authors should report the variance across folds and, ideally, a paired significance test, to support the claim that the improvement is systematic rather than noise.","section":"Section V-C / Section V-D, Tables I and III"},{"comment":"The paper does not specify how the joint histogram used in Eq. (6) is computed in practice: the number of bins, the bin boundaries, and whether the sum runs over all (i,j) pairs (as written) or over corresponding positions (A2D(i), A3D(i)) are all unspecified. If the implementation actually used paired indices, then the loss is different from Eq. (6) and the correct formula must be given. As written, the loss cannot be reproduced, and the discrepancy between the equation and any plausible implementation is material.","section":"Section IV-E and Section V-C"}],"minor_comments":[{"comment":"The word 'pase' should be 'pose.'","section":"Section V-E, first sentence"},{"comment":"The entry 'Posteret al.' is missing a space, and the footnote marker '1The method...' is formatted inconsistently, making the table hard to read.","section":"Table III"},{"comment":"The notation 'G VB0' and 'P VB0' for ARL-VTF gallery/probe designations is not defined in the text or in a reference to the dataset's protocol, which reduces reproducibility.","section":"Section V-A"},{"comment":"The text says 'Without the Joint Entropy minimization, Ours + JAM showcased the highest performance ... on FaceScape,' but Table I shows that Ours+JAM+JE has a higher 90+ TAR@1% FAR (78.910 vs. 78.501). This should be stated more precisely to avoid an apparent contradiction.","section":"Section V-D"}],"recommendation":"reject","confidential_remarks":"The central mechanism in Eq. (6) is mathematically incorrect as written: the 'joint entropy' is a sum of marginal entropies, so the loss cannot promote cross-modal attention consistency. This is a load-bearing error for the paper's main contribution. A corrected paired-histogram loss would be a different method and would require new experiments; the current manuscript cannot be fixed by minor edits. The empirical gains are plausible and the JAM idea may be worth pursuing, but the paper's central claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a real empirical contribution: a 2D-3D domain-adaptation framework for pose-robust face recognition, combining shared query/key attention (JAM) with an entropy regularizer, evaluated on FaceScape and cross-dataset on ARL-VTF. The gains are substantial—7.8% better TAR@1%FAR than CrossPoint on 90+ profile views, and the cross-dataset generalization is genuinely useful for surveillance-style pipelines. The ablations are thorough, and the authors are honest about the need for aligned 2D-3D training data. That part deserves credit.\n\nThe problem is the joint entropy loss. Equation 6 defines P(a2D,a3D) as a normalized sum over all i,j pairs of indicator functions on discretized attention values. That sum factorizes into the product of the two marginal histograms. So H(A2D,A3D) = H(A2D) + H(A3D). Minimizing it does not align corresponding spatial locations in the two attention maps; it just makes each map more peaked on its own. The paper's repeated claim that this loss \"enhances correlations among intersecting 2D and 3D representations\" is not supported by the math. The stated mechanism in Sections I, III, and IV-E is internally inconsistent with Eq. 6.\n\nThat is a load-bearing flaw, but it is fixable. The authors likely intended a paired histogram over corresponding elements (i.e., summing over the diagonal i=j), which would give a true joint distribution and a real consistency loss. Alternatively, they could reframe the contribution as marginal entropy regularization on top of JAM, though then the novelty shrinks. As written, the abstract's causal story is wrong.\n\nMissing details make it worse: no code, no binning parameters for the histograms, and no error bars in the tables (only the ROC curves have shaded regions). The hyperparameter tuning of AdaFace margins is standard and not circular, so the reader's concern there is misplaced.\n\nI would send this to peer review, but with a strong request to correct or reframe the loss. The empirical results are too good to desk-reject outright; the method likely works for reasons not stated in the paper. If the authors can fix the loss or honestly reframe the contribution, this could be a solid applied paper. As it stands, the central claim is unsupported. I'd bring it to a reading group as a cautionary example of how a subtle equation error can undermine an otherwise careful empirical study.","headline":"A useful empirical recipe undercut by a mis-specified joint entropy loss; the paper's central mechanism does not do what it claims, though the experiments are solid enough to warrant a serious look if the authors can fix the math.","tokens_in":18626,"tokens_out":2381,"would_cite":false,"duration_ms":25797,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By training a 2D face network with a shared 2D-3D attention mapping and a joint entropy loss, the paper reports profile-view (90°+) TAR@1%FAR gains of at least 7.1 points on FaceScape and 1.57 points on ARL-VTF while keeping inference…","keywords":["2D-3D domain adaptation","pose-invariant face recognition","joint attention mapping","joint entropy regularization","point cloud priors","profile face recognition","TAR@1%FAR"],"falsifier":"On a trained model, compute the paper's $L_{JE}(A_{2D}, A_{3D})$ and compare it to $H(A_{2D}) + H(A_{3D})$; equality for every batch would show the joint entropy carries no location-pairing information, and any pose-invariance gain would then have to come from the shared attention mapping alone. A direct check is to retrain with the JE term replaced by that marginal-entropy sum and see whether the ARL-VTF profile TAR@1%FAR gain disappears.","tokens_in":17505,"feed_emoji":"👤","tokens_out":9544,"duration_ms":83982,"temperature":0.7,"pith_summary":"The paper aims to make ordinary 2D face recognition hold up under large pose differences by letting the 2D network, during training only, learn from pose-invariant 3D point-cloud representations. Its two new pieces are a shared 2D-3D joint attention mapping (JAM), which computes attention for both modalities with the same filters, and a joint entropy (JE) loss, which is intended to keep the two attention maps consistent. Evaluated on FaceScape, the combined framework reaches 78.910% TAR@1%FAR at 90°+ profile pose, at least 7.1 percentage points above the compared domain-adaptation baselines; on the 2D-only ARL-VTF benchmark it improves profile TAR@1%FAR by 1.57 points over the best comparison. If the result holds, deployed face recognition could gain pose invariance without needing 3D sensors, 3D enrollment data, or frontalization at inference time.","feed_headline":"Profile face recognition gains 7.1% via 3D priors","feed_subtitle":"Shared attention plus entropy training transfers 3D pose invariance into a 2D-only inference pipeline.","key_machinery":"The load-bearing mechanism is the pair (JAM, JE). JAM is the attention operation $J_d = \\gamma\\,\\mathrm{softmax}(Q(z_d)K^\\top(z_d))V_d(z_d) + z_d$, where $Q$ and $K$ are 1x1 convolutions shared between the 2D and 3D branches, so both modalities are forced to attend with the same filters, while the value projections $V_d$ remain modality-specific. JE is the joint entropy $H(A_{2D}, A_{3D})$ of the discretized attention maps, computed from a normalized histogram over pairs of attention values and minimized together with the two identification losses. The paper's stated mechanism is that minimizing this entropy regularizes the shared attention so that consistent patterns across the two maps are emphasized and pose-variant 2D patterns are de-emphasized, which is what the ablation isolates when JE is removed.","core_discovery":"The central claim is that pose invariance can be transferred from a 3D point-cloud network to a 2D face embedding network through a shared attention mechanism, without ever requiring 3D input at inference. The framework passes 2D and 3D feature maps through identical query and key projections, so the resulting attention maps emphasize the same spatial and channel patterns in both domains, and it adds a joint entropy regularizer computed on the discretized attention maps to penalize inconsistency between them. With this design, the paper reports TAR@1%FAR of 78.910% for 90°+ views on FaceScape, a gap of only 18.19 percentage points from frontal performance, and a profile TAR@1%FAR of 54.678% on ARL-VTF after training only on FaceScape, which beats the compared methods despite a 2.605-point drop on frontal ARL-VTF views that the paper attributes to domain shift.","pith_inferences":["Editorial extension: if the histogram in Eq. 6 sums over all location pairs, the joint entropy equals the sum of the two marginal entropies, so the regularizer would encourage each attention map to be concentrated rather than aligning corresponding 2D and 3D locations; this can be tested by replacing $L_{JE}$ with $H(A_{2D})+H(A_{3D})$ and re-running the reported protocol.","Editorial extension: a location-paired loss such as mutual information between corresponding attention positions would directly implement the paper's consistency story and could show whether location-wise alignment, rather than sparsity, drives the profile-view gains.","Editorial extension: the recipe of shared attention projections plus an entropy regularizer could generalize to other pose-sensitive 2D tasks that have a 3D morphable model, such as head-pose estimation or expression recognition, although the paper does not claim this."],"forward_implications":["Deployed 2D face recognition gains pose invariance without any 3D capture at inference time, because 3D data is used only during training.","Models trained on a paired 2D-3D dataset transfer to 2D-only operational benchmarks: the framework's ARL-VTF profile TAR@1%FAR is 54.678%, above every compared method, with average TAR@1%FAR of 68.550%.","The frontal-to-profile performance gap narrows to 18.19 percentage points on FaceScape, versus 23.78 for the strongest comparison, meaning the benefit is concentrated at the extreme poses that break conventional systems.","Because inference needs only a 2D image, the approach can be inserted into existing recognition pipelines without keypoint detection, frontalization, or synthetic-view generation.","The main practical limit is the stated requirement for an aligned 2D-3D training corpus; the method inherits the cost and availability constraints of such datasets."],"supporting_citations":[{"why":"Supplies the cross-modal contrastive baseline whose profile-view performance the FaceScape comparison is built against.","marker":"[1]"},{"why":"Supplies the MMD regularizer baseline used in both FaceScape and ARL-VTF comparisons.","marker":"[21]"},{"why":"Supplies the CORAL alignment baseline used as a domain-adaptation comparison.","marker":"[60]"},{"why":"Supplies the RST cross-domain identification baseline that the framework is compared against on both datasets.","marker":"[18]"},{"why":"Supplies the 3D backbone that provides the pose-invariant point-cloud features used during training.","marker":"[55]"},{"why":"Supplies the 2D backbone whose representations are adapted toward the 3D domain.","marker":"[56]"},{"why":"Supplies the quality-adaptive margin classifier and loss used for identification in both domains.","marker":"[35]"},{"why":"Provides the paired 2D image and 3D mesh dataset used for training and the FaceScape evaluation.","marker":"[74]"},{"why":"Provides the 2D-only ARL-VTF dataset used to test cross-dataset generalization to profile poses.","marker":"[53]"}],"fun_headline_variants":["2D face ID gets 3D pose invariance","Shared attention transfers 3D pose invariance to 2D face net","Joint entropy and attention improve pose-robust face ID","Profile TAR gains 7.1% with 3D-prior attention","Cut pose gap in face ID via 2D-3D attention"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the joint entropy of the two attention maps, as built from a histogram over all pairs of discretized attention values, actually couples corresponding 2D and 3D attention locations; if that histogram instead treats the two maps as independent, the regularizer cannot enforce cross-modal consistency by itself.","fun_headline_variants_meta":{"raw":{"variants":["2D face ID gets 3D pose invariance","Shared attention transfers 3D pose invariance to 2D face net","Joint entropy and attention improve pose-robust face ID","Profile TAR gains 7.1% with 3D-prior attention","Cut pose gap in face ID via 2D-3D attention"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00093,"raw_usage":{"total_tokens":3990,"prompt_tokens":960,"completion_tokens":3030,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":2939}},"tokens_in":576,"tokens_out":3030,"duration_ms":20414,"temperature":1.0,"reasoning_tokens":2939,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:40:33.675057+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a trained model, compute the paper's $L_{JE}(A_{2D}, A_{3D})$ and compare it to $H(A_{2D}) + H(A_{3D})$; equality for every batch would show the joint entropy carries no location-pairing information, and any pose-invariance gain would then have to come from the shared attention mapping alone. A direct check is to retrain with the JE term replaced by that marginal-entropy sum and see whether the ARL-VTF profile TAR@1%FAR gain disappears.","supporting_citations":[{"cited_title":"Gretton, K","cited_arxiv_id":null,"evidence_quote":"Supplies the MMD regularizer baseline used in both FaceScape and ARL-VTF comparisons."},{"cited_title":"Sun and K","cited_arxiv_id":null,"evidence_quote":"Supplies the CORAL alignment baseline used as a domain-adaptation comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RST cross-domain identification baseline that the framework is compared against on both datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 2D backbone whose representations are adapted toward the 3D domain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the paired 2D image and 3D mesh dataset used for training and the FaceScape evaluation."},{"cited_title":"Poster, M","cited_arxiv_id":null,"evidence_quote":"Provides the 2D-only ARL-VTF dataset used to test cross-dataset generalization to profile poses."}],"review_version":1}