{"id":"17075d59-218f-493d-aea0-fd3684638353","arxiv_id":"2412.19849","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"A single-image face reconstruction pipeline that removes eyeglasses by face parsing and inpainting, then adds a learned bump map to a 3D morphable model for geometry details.","lead":"This paper proposes a two-step pipeline for 3D face reconstruction from a single photo. The pipeline first removes eyeglasses from the image using face parsing and an inpainting network, then regresses a 3D morphable model and adds a learned bump map for mid-level geometry details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The geometry-detail loss L_geo (Eq. 6) has no described source for ground-truth bump maps Φ̃ and no described network predicting Φ, leaving the paper's core claim without a defined training signal.","rationale":"The reader's weakest assumption points to the undefined ground-truth bump map Φ̃ in Eq. (6), which is indeed the most immediate defect. My read agrees and slightly broadens it: the estimated bump map Φ also has no described predictor, and the training protocol for L_geo (how it interacts with L_shape, what data provides supervision) is entirely absent. Since the paper's central claim is precisely the addition of geometry details via bump maps, and the mechanism for learning them is undefined, the claim cannot be verified from the manuscript. This confirms the reader's REJECT verdict. I mark agreement as partial rather than full because the reader focused on the missing Φ̃, while I see the missing Φ-predictor and missing integration of L_geo as equally load-bearing; both stem from the same omission of a detailed-geometry supervision pipeline. A CONDITIONAL verdict would be possible if the authors supplied the missing supervision details, but as written the manuscript does not make that content available.","tokens_in":8465,"tokens_out":3662,"duration_ms":35081,"concrete_test":"Audit the four cited training sources (Helen, CelebAMask-HQ, 300W, AFLW) for any 3D ground-truth geometry or depth data; none provides bump maps. Then attempt to compute the gradient of Eq. (6) on these datasets: without a defined Φ̃ and a defined predictor for Φ, the loss is undefined. If the authors instead obtain Φ̃ from a 3DMM or rendered scans, that procedure is absent from Section 4 and must be specified before the geometry-detail claim can be evaluated. A positive result would be finding an explicit description of the bump-map supervision pipeline in a supplement or a released training code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central novelty is adding mid-level geometry details via bump maps. Section 3.3 defines the bump map loss in Eq. (6) as L_geo = ||Φ̃ − Φ|| + ||∂Φ̃/∂x − ∂Φ/∂x|| + ||∂Φ̃/∂y − ∂Φ/∂y||, where Φ̃ is called the ground truth. The paper never states (a) how the estimated Φ is produced from the regressed coefficient vector y ∈ R^239, (b) where Φ̃ comes from, or (c) how L_geo is combined with L_shape in the training objective. The datasets listed in Section 4 (Helen, CelebAMask-HQ, 300W, AFLW) are 2D landmark and segmentation datasets; none contains depth scans or bump maps. Without a defined ground-truth bump map, the gradient of Eq. (6) cannot be computed, so the claimed detail generation cannot be trained. This is not a disagreement with any baseline or consensus; it is an internal omission. The central claim therefore rests on an unspecified, likely unavailable supervision signal, making the method irreproducible and the headline result unfalsifiable as described.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a two-stage pipeline for single-image 3D face reconstruction under occlusion. In the first stage, a face parsing network and a face edge-line network produce a parsing map and an edge map, which are combined using the authors' earlier algorithm [42] and fed with the input image to a face image synthesis network that removes eyeglasses. In the second stage, a ResNet-50 regresses a 239-dimensional vector of 3DMM-related coefficients, and a bump map is introduced to add mid-level geometry detail. The loss in Eq. (6) is meant to supervise this bump-map detail. Experiments consist of qualitative comparisons with Sela et al., PRNet, and 3DDFA, plus quantitative LFW face-verification tests of reconstructed shapes with and without eyeglasses.","tokens_in":8762,"tokens_out":6448,"duration_ms":56919,"significance":"If the proposed separation of global shape regression from bump-map detail prediction under occlusion were fully specified and validated, it would address a real gap: most single-image 3DMM methods either ignore mid-level geometric details or degrade under occlusion. The authors provide concrete weight values for the shape loss, compare against three baselines qualitatively, and report numerical LFW verification results rather than showing only subjective images. However, the central training signal for the claimed novelty is undefined, the estimated bump map has no described network or integration path, and the only quantitative table shows the proposed method below the cited baseline. No code or data are provided, so the manuscript as written is not reproducible and the headline contribution is not falsifiable.","major_comments":[{"comment":"The geometry-detail loss L_geo is defined as ||Φ̃ − Φ|| + ||∂Φ̃/∂x − ∂Φ/∂x|| + ||∂Φ̃/∂y − ∂Φ/∂y||, where Φ̃ is called the ground truth. The manuscript never states where Φ̃ comes from. The datasets listed in Section 4 (Helen, CelebAMask-HQ, 300W, AFLW) are 2D image, landmark, and segmentation datasets; none provides depth scans, displacement maps, or bump maps. Without a defined source of Φ̃, the gradient of Eq. (6) cannot be computed, so the claimed training of geometry details is undefined. This is an internal omission, not a disagreement with a baseline, and it makes the paper's central claim unfalsifiable as written.","section":"Section 3.3, Eq. (6)"},{"comment":"The definition of the bump map is tautological: Φ(b) is defined as φ(d′(b) − d(b)) and then inverted to obtain d′(b) = d(b) + φ^{-1}(Φ(b)). The text never describes a network or algorithm that predicts Φ from the regressed vector y ∈ R^239, nor how Φ is combined with the base shape S_base to produce the final detailed mesh. The phrase 'Inspired by the method of image-to-image translation' is not an implementation. Thus even if a ground-truth Φ̃ were available, the estimated Φ is unspecified and the pipeline cannot be reproduced.","section":"Section 3.3, Eq. (5)"},{"comment":"The quantitative evidence contradicts the claimed robustness. For LFW verification, the proposed method with glasses gives EER 84.37 ± 1.44, accuracy 85.79 ± 0.42, and nAUC 92.87 ± 1.09, all lower than Tran et al.'s 89.40 ± 1.52, 89.36 ± 1.25, and 95.90 ± 0.95. Even the no-glasses result (nAUC 95.37 ± 0.65) is slightly below Tran et al. The text states that the limited drop 'demonstrating the robustness of our method,' but the numbers show the reconstruction is less discriminative than the baseline, so the quantitative claims of state-of-the-art robustness are not supported by the presented evidence.","section":"Table 1"},{"comment":"The final face parsing map M_fin is obtained following 'Zhao et al.'s Algorithm [42],' which is the authors' own prior work. The algorithm is not summarized or independently validated in this manuscript. Because the eyeglass-removal step and the subsequent reconstruction depend on M_fin, a reader cannot evaluate whether the proposed pipeline, as opposed to the prior work, is responsible for the occlusion handling. At minimum, the dependency should be described or the component ablated.","section":"Sections 3.1, 3.3, and reference [42]"}],"minor_comments":[{"comment":"The text says the ResNet 'was trained under weakly supervised' and later says 'we used a fixed ResNet-50 network to regress these coefficients'; please clarify whether the regressor is fixed after a separate training stage or jointly optimized.","section":"Section 3.3"},{"comment":"The symbol d_gt is used both as a scalar 0/1 condition and as a 'ground-truth distance matrix'; please separate the thresholding condition from the distance map notation.","section":"Eq. (1)"},{"comment":"The sentence 'we found that by adding these last two terms ... reduce bump map noise' is presented without experimental support; an ablation of the two gradient terms in Eq. (6) should be added.","section":"Section 3.3"},{"comment":"The number of U-Net stacks M and the message-passing area partition are never specified; please provide concrete architecture details so the method can be reimplemented.","section":"Section 3.2"},{"comment":"There are several presentation issues: 'Vetoer's first 3D face' should be 'Vetter'; reference [31] is attributed to 'Gusi et al.' in Section 3.1 but to 'Te et al.' in Section 4; and the abstract's claim of 'state-of-the-art qualitative performance' should be reconciled with the less favorable quantitative results in Table 1.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The manuscript relies heavily on the authors' own prior work [42] for the core occlusion-handling step without describing it, and the central geometry-detail contribution lacks both a defined supervision signal and a defined prediction network. The one quantitative comparison is unfavorable to the proposed method. These issues are load-bearing and would require, at minimum, a substantially rewritten method section with new training-data definitions and a valid quantitative evaluation; as written, the paper is not reproducible. I do not see a path to acceptance within the current scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: the paper's headline claim — detailed 3D face reconstruction under occlusion via bump maps — is not evaluable from the text, because the training signal for the bump map is missing. The reader's rejection is sound, and the stress-test concern is real.\n\nWhat is genuinely there: a coherent pipeline that combines face parsing, edge-line generation with message passing, face inpainting, and a 3DMM regressor. The idea of first removing eyeglasses via inpainting before fitting the 3D shape is sensible and could be useful in practice. The bump-map loss in Eq. (6), with the two gradient terms, is a small but reasonable extension to encourage smooth yet detailed surfaces. If the details were fully specified, this would be an incremental but usable contribution.\n\nThe problems are load-bearing. Eq. (6) requires ground-truth bump maps Φ̃, but the paper never says where they come from. The datasets listed in Section 4 — Helen, CelebAMask-HQ, 300W, AFLW — are 2D parsing/landmark sets; none contain depth or displacement maps. Without a source for Φ̃, L_geo cannot be computed and the 'geometry detail' claim is untestable. Eq. (5) defines the bump map as an encoding of the depth difference and then inverts it, so the output is a fitted residual on the 3DMM, not a separately learned mid-level representation. The first contribution claims a novel Face Image Synthesis Network, but Section 3.3 says I_out is synthesized 'by existing methods [15]' (MaskGAN). That is a direct internal contradiction. The quantitative evaluation is also weak: Table 1 shows the method underperforms Tran et al., and there is no quantitative comparison against the reconstruction baselines shown in Figure 4. The abstract's mention of hair and palms is not supported anywhere — the method only addresses eyeglasses.\n\nNone of this requires figures we can't see. The text-level omissions are enough to reject the current formulation. I would not send this to reviewers; they'd spend the time rediscovering the missing ground truth. If the authors have a real training set for Φ̃ and a description of the synthesis network, they should rewrite the paper with those details.\n\nRecommendation: desk reject.","headline":"Load-bearing gaps in the bump-map training signal and a contradiction about the synthesis network make the core claim unevaluable; the pipeline is coherent but not referee-ready.","tokens_in":9239,"tokens_out":3287,"would_cite":false,"duration_ms":30099,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single color photo of an occluded face becomes a detailed 3D shape by removing the occluder in 2D, then adding a learned bump map to a coarse morphable model.","keywords":["3D face reconstruction","bump mapping","face parsing","occluded scenes","face image synthesis","geometry details","single image reconstruction","face edge lines"],"falsifier":"Run the pipeline on a set of photos whose ground-truth high-resolution 3D face scans are known, and measure the distance between the reconstructed surface and the scan around the eyes; if the bump-mapped shape is no closer to the scan than the bare morphable-model shape, or if no ground-truth bump map can be obtained to train Equation (6), the central detail-recovery claim fails.","tokens_in":8273,"feed_emoji":"👓","tokens_out":8924,"duration_ms":74575,"temperature":0.7,"pith_summary":"Single-image 3D face reconstruction normally produces smooth shapes and strong textures, and it fails when glasses, hair, or hands cover part of the face. This paper proposes a two-step pipeline: first it removes the occluder in 2D by combining face parsing and edge-line maps, synthesizing an unobstructed frontal image; then it regresses a coarse 3D morphable shape and adds mid-level surface detail with a learned bump map. The central claim is that this unified design recovers plausible geometry details and stable global shapes from occluded real-world photos, where texture-focused baselines break down. If the claim holds, single-image reconstruction can preserve wrinkles and expression folds without multiview or depth input, even under occlusion.","feed_headline":"Bump maps add mid-level detail to 3D faces from single photos","feed_subtitle":"A two-stage pipeline first removes eyeglasses, then adds wrinkles and folds to the coarse shape.","key_machinery":"The load-bearing mechanism is the bump map $\\Phi$ added to the base 3D morphable model: a grayscale depth-displacement field that lets the pipeline add mid-level wrinkles and folds without regressing high-resolution geometry directly. Around it sit two supporting modules: a face parsing map generator with edge-aware graph reasoning, and a face edge-line generator built from stacked encoder-decoder networks with message passing between visible and eyeglass-covered regions plus an adversarial effectiveness discriminator. These maps feed a face image synthesis step that regenerates the unobstructed frontal image, and the final parameter vector $y = (\\alpha_{id}, \\beta_{exp}, \\beta_t, \\gamma, p) \\in \\mathbb{R}^{239}$ covers identity, expression, texture, illumination, and pose while detail is delegated to the bump map.","core_discovery":"The paper claims that mid-level facial geometry details can be recovered from a single color image even when parts of the face are occluded, by separating a stable coarse shape estimate from a detail estimate. The coarse shape comes from a weakly supervised ResNet regressor applied to a synthesized glasses-free frontal image; the detail comes from a bump map $\\Phi$ that encodes per-pixel depth displacement relative to the base shape, so the detailed depth is $d'(b) = d(b) + \\phi^{-1}(\\Phi(b))$. The training loss for this detail stage combines an L1 term on the bump map with L1 terms on its 2D gradients, which the authors report suppresses noise while preserving high-frequency detail. Their stated result is high-quality reconstruction with geometry details from captured facial images under occluded scenes, with the eyeglass-removal stage as the key to making the global shape regression stable.","pith_inferences":["Editorial inference: the quality ceiling of the whole pipeline is set by the 2D eyeglass-removal stage, so failures in face synthesis are likely to dominate errors in occluded regions regardless of the 3D regressor.","Editorial inference: the geometry-detail component is not independently reproducible until the authors disclose the source of the ground-truth bump maps used in the detail loss.","Editorial inference: the coarse-shape/detail decomposition suggests a drop-in extension in which any single-image shape regressor could be paired with the same bump-map refinement network.","Editorial inference: the reported verification-drop measure is an indirect proxy; a direct comparison against high-resolution 3D scans would be a stronger test of whether the added details are geometrically accurate."],"forward_implications":["Reconstructed faces from eyeglass-occluded photos should show visible wrinkles and folds instead of a smooth, glasses-flattened surface.","The same pipeline can handle multiple occlusion types such as glasses, hair, palms, food, and fingers simultaneously, because the 2D synthesis step regenerates a clean frontal face before shape regression.","Face verification on the reconstructed shapes degrades only slightly when glasses are present, indicating the occluder no longer corrupts the recovered identity shape.","Because coarse shape and detail are estimated separately, an error in local detail estimation does not destabilize the global face shape.","The bump-map representation keeps the added detail inexpensive, avoiding a full high-resolution mesh regression on every pixel."],"supporting_citations":[{"why":"Supplies the parametric morphable model whose coarse shape is the base that the bump map refines.","marker":"[2]"},{"why":"Provides the weakly supervised shape-coefficient regression used to estimate the coarse 3D face from the synthesized unobstructed image.","marker":"[5]"},{"why":"Supplies the face image synthesis network that regenerates the frontal face after eyeglass removal, the input to the 3D regressor.","marker":"[15]"},{"why":"Gives the algorithm that fuses the parsing map and edge-line map into the final face parsing map used for occlusion handling.","marker":"[42]"},{"why":"Supplies the edge-aware face parsing backbone and graph reasoning module that produce the initial face parsing map.","marker":"[31]"},{"why":"Provides the unrestricted facial geometry reconstruction baseline used to demonstrate the advantage under eyeglass occlusion.","marker":"[29]"},{"why":"Provides the position-map regression baseline compared in occluded scenes.","marker":"[6]"},{"why":"Provides the dense 3D face alignment baseline compared in occluded scenes.","marker":"[7]"}],"fun_headline_variants":["Bump maps add mid-level detail to occluded 3D faces","Single-photo 3D faces get geometry details despite hair, hands, glasses","Two-stage pipeline recovers depth details even with facial occlusions","Bump mapping plus glasses removal yields detailed 3D faces from one shot","Occlusion-aware face reconstruction adds wrinkles and folds to coarse shapes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The geometry-detail loss in Equation (6) assumes a ground-truth bump map $\\tilde{\\Phi}$ is available for training, but the paper never says where that map comes from; without that supervision, the claim of adding geometry details cannot even be evaluated.","fun_headline_variants_meta":{"raw":{"variants":["Bump maps add mid-level detail to occluded 3D faces","Single-photo 3D faces get geometry details despite hair, hands, glasses","Two-stage pipeline recovers depth details even with facial occlusions","Bump mapping plus glasses removal yields detailed 3D faces from one shot","Occlusion-aware face reconstruction adds wrinkles and folds to coarse shapes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001036,"raw_usage":{"total_tokens":4306,"prompt_tokens":836,"completion_tokens":3470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":3374}},"tokens_in":452,"tokens_out":3470,"duration_ms":24200,"temperature":1.0,"reasoning_tokens":3374,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:20:17.585942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a set of photos whose ground-truth high-resolution 3D face scans are known, and measure the distance between the reconstructed surface and the scan around the eyes; if the bump-mapped shape is no closer to the scan than the bare morphable-model shape, or if no ground-truth bump map can be obtained to train Equation (6), the central detail-recovery claim fails.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops","cited_arxiv_id":null,"evidence_quote":"Provides the weakly supervised shape-coefficient regression used to estimate the coarse 3D face from the synthesized unobstructed image."},{"cited_title":"In: Computer Graphics International Conference","cited_arxiv_id":null,"evidence_quote":"Gives the algorithm that fuses the parsing map and edge-line map into the final face parsing map used for occlusion handling."},{"cited_title":"In: European Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Supplies the edge-aware face parsing backbone and graph reasoning module that produce the initial face parsing map."}],"review_version":1}