{"id":"2b1c9270-7025-47d4-aa10-8361b86461de","arxiv_id":"2505.06038","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A multitask fusion network with inter-task feature aggregation and gating reports state-of-the-art document dewarping results on DIR300, DocUNet, and DocReal, subject to comparison caveats.","lead":"The authors propose SalmRec, a neural network that flattens photos of curled or folded documents by jointly predicting the paper's 3D shape, its flat coordinates, and horizontal and vertical line directions, then fusing these features with gating. It reports top scores on three document dewarping benchmarks, but the gains over prior work are modest and the comparisons are not all made under identical training conditions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"On DocReal, Table 2 contradicts the unqualified SOTA claim: Ours is not best on ED/CER.","rationale":"The reader correctly identified the uncontrolled baseline comparison as a threat to the SOTA claim, but the more decisive problem is internal to the paper: Table 2 on DocReal shows Ours is not state-of-the-art on ED/CER. This contradiction does not depend on how baselines were produced; it is visible in the authors’ own numbers. The paper can likely be repaired by qualifying 'state-of-the-art' to geometric metrics (MS-SSIM, LD, AD), or by adding the missing caveat about OCR performance on DocReal. Because the geometric metrics do support SOTA on all three benchmarks, and the ablations provide credible evidence for the proposed modules, the appropriate decision remains conditional acceptance pending revision of the overbroad claim and fair, fully reported comparisons. The reader’s identified weak assumption, while valid, is a separate concern about reproducibility and comparability rather than an internal inconsistency.","tokens_in":16220,"tokens_out":7229,"duration_ms":68938,"concrete_test":"Perform a per-metric ranking on DocReal Table 2. If any baseline has a lower ED or CER than Ours (which the table already shows: PaperEdge and DocReal), then the unqualified claim 'state-of-the-art on three public benchmarks' is false as written. To rule out an evaluation artifact, rerun the official DocReal OCR metric code on the released outputs of Ours, PaperEdge, and DocReal; if Ours becomes strictly best on all five metrics, the claim can be restored.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper’s stated central claim—'Our method achieves state-of-the-art performance on three public benchmarks' (Contributions, §1)—is contradicted by its own Table 2 on the DocReal benchmark. In that table, Ours reports ED 193.06 and CER 0.2604, while PaperEdge reports ED 186.86 and CER 0.2451, and DocReal reports ED 184.54 and CER 0.2485. Thus on two of the five reported metrics, Ours is not the best method. The DocReal section (§4.4) only claims improvements over DocReal on MS-SSIM, LD, and AD, and the comparison to text-line methods is also incomplete: RDGR’s ED (188.00) is lower than Ours (193.06) even though Ours has a better CER (0.2604 vs 0.2768). If the authors intend 'state-of-the-art' to apply only to geometric metrics, that qualification is absent from the abstract and the contribution bullet, both of which also emphasize text recognition accuracy. This is an internal inconsistency in the paper’s own reported numbers, independent of whether the baseline evaluations were fair. The reader’s concern about uncontrolled re-evaluation of baselines remains valid, but it is a secondary issue for the DocReal SOTA statement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SalmRec, a multitask document image rectification network that predicts a 2D deformation field with four auxiliary tasks: 3D coordinate maps, UV maps, horizontal lines, and vertical lines. The network includes an inter-task feature aggregation module (FA) that reweights each task's features using a leave-one-out global attention mechanism, and a gating module (Gate) that balances global and local task features before a Transformer decoder predicts the deformation field. The model is trained on the synthetic DocDewarpHV1 dataset and evaluated on three benchmarks: DIR300, DocReal, and DocUNet. The paper reports state-of-the-art results on all three benchmarks and presents ablations showing that the full model with FA and Gate outperforms the four-task baseline without these modules.","tokens_in":16481,"tokens_out":3765,"duration_ms":37535,"significance":"The paper addresses a real and current problem in document dewarping: negative interference among auxiliary tasks in multitask rectification networks. The proposed FA and Gate modules are simple, interpretable, and the authors' self-contained ablation (Tables 4 and 5) consistently shows that the full model improves over the four-task baseline on all three benchmarks, giving internal support for the architecture. If the claimed results hold under controlled comparison, the work would be a useful contribution to the document image rectification literature. However, the external significance is substantially weakened by two issues: the headline state-of-the-art claim is internally contradicted by the authors' own Table 2 on DocReal, and the main comparisons are confounded by training on a different synthetic dataset (DocDewarpHV1) than most baselines (Doc3D). The paper does not provide code or trained models, and the absence of a controlled retraining experiment makes the cross-dataset comparison difficult to interpret.","major_comments":[{"comment":"The unqualified claim in the introduction and abstract that the method 'achieves state-of-the-art performance on three public benchmarks' is contradicted by Table 2 on DocReal. In that table, Ours reports ED 193.06 and CER 0.2604, while PaperEdge reports ED 186.86 and CER 0.2451, and DocReal reports ED 184.54 and CER 0.2485. On two of the five reported metrics, Ours is not the best. The text in §4.4 correctly limits the improvement over DocReal to MS-SSIM, LD, and AD, but the abstract and contribution bullet emphasize text recognition accuracy as part of the state-of-the-art claim. This internal inconsistency must be resolved, either by qualifying the SOTA claim to geometric metrics or by providing a reasoned justification for why the OCR results on DocReal should not be taken at face value.","section":"§1 and §4.4 (Table 2)"},{"comment":"The state-of-the-art comparisons are confounded by a training-data mismatch. The model is trained on DocDewarpHV1 (Section 4.1), while most of the compared baselines were trained on Doc3D. Different synthetic training data can change rectification performance substantially, especially for OCR metrics, so the reported improvements may partly reflect the training distribution rather than the proposed architecture. The manuscript notes that Doc3D lacks fine-grained annotations and that prior methods use binarization or line-detection models, which is a legitimate motivation for using DocDewarpHV1. However, the authors then compare their numbers to published results from models trained on Doc3D without a controlled experiment. To support the state-of-the-art claim, the authors should retrain, or otherwise evaluate, baselines on the same training set, or carefully discuss the expected effect of the training-data difference and provide controlled ablations to separate the contribution of the architecture from the contribution of the training data.","section":"§4.1 and Tables 1–3"},{"comment":"The ablation of the individual FA and Gate modules in Table 5 shows mixed and sometimes negative effects on the DocUNet benchmark: adding FA alone (row 2) worsens LD from 7.28 to 7.30 and ED from 350.5 to 376.8 relative to the four-task baseline, and adding Gate alone (row 3) worsens MS-SSIM from 0.50 to 0.49 and LD from 7.28 to 7.49. The full model (row 4) improves over the baseline on all three benchmarks, and the paper states that the two modules jointly serve the task. This is a defensible position, but the current presentation does not discuss why a module that alone can degrade performance is still beneficial when combined. The authors should address this apparent negative interference explicitly, for example by analyzing the learned gate weights or by reporting per-benchmark variance, to strengthen the claim that the combination is robust and not a result of fortuitous hyperparameter choices.","section":"§4.5, Table 5"},{"comment":"The OCR evaluation uses two different engines (Tesseract for English, PaddleOCR for Chinese) and different image subsets for the two English benchmarks (90 images for DIR300, '50 and 60 images' for DocUNet). The sentence in Section 4.3 saying 'the text recognition accuracy is evaluated on 50 and 60 images, respectively' is ambiguous when read with Table 3, which shows two ED/CER numbers per method. It appears that the first number in each pair corresponds to a 50-image subset and the second to a 60-image subset, but this is not stated explicitly in the table or its caption. Because the SOTA claim depends on the comparability of these numbers, the authors must clarify the exact evaluation protocol and ensure that all methods are evaluated on the same image subsets with the same OCR engine and preprocessing.","section":"§4.3, OCR evaluation protocol"}],"minor_comments":[{"comment":"The word 'Bases' in the title should be 'Based'.","section":"Title"},{"comment":"The caption reads 'Predcit results of our proposed method'; 'Predcit' should be 'Prediction'.","section":"Figure 6 caption"},{"comment":"The method name 'Pice-Wise' should be 'Piece-Wise' in both the table and the text in Section 4.4.","section":"Table 3"},{"comment":"In the paragraph discussing Table 5, the sentence 'The Character Error Rate (CER) reached an impressive 10.3%.' is unclear because it is not stated whether 10.3% is a relative or absolute improvement. Please specify the reference baseline.","section":"Section 4.5"},{"comment":"The sentence 'On the DocUNet benchmark, the text recognition accuracy is evaluated on 50 and 60 images, respectively' could be interpreted as two separate evaluations; please clarify the exact subsets and whether the results are reported as pairs in Table 3.","section":"Section 4.2"},{"comment":"The word 'benchamrks' appears in the ablation discussion; it should be 'benchmarks'.","section":"Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be a preprint that may have undergone minimal copyediting, given the title typo and the repeated formatting errors. The central architectural contribution is plausible and the internal ablation is the strongest part of the manuscript, but the state-of-the-art claim needs to be substantially revised and the evaluation made more controlled before the paper can be considered for publication. I recommend that the editor require the authors to address the DocReal inconsistency and the training-data confound in their revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thoughts on arXiv:2505.06038 (SalmRec).\n\nBottom line: this is a credible empirical architecture paper with an honest self-ablation, and an overstated headline. The two proposed modules — the leave-one-out inter-task feature aggregation and the gating between global/local features — are a genuinely new combination in the document-rectification subfield, and Tables 4 and 5 back them up. The all-four-tasks row without FA/Gate does worse on OCR than using UV alone (CER 0.1646 vs 0.1556 on DIR300), which is a good motivating observation: the interference is real and the modules measurably reduce it. That part is controlled and internally consistent.\n\nThe soft spot is the SOTA claim. The contribution bullet says \"state-of-the-art on three public benchmarks,\" but the paper's own Table 2 contradicts that on DocReal: on ED and CER, both PaperEdge (186.86/0.2451) and DocReal (184.54/0.2485) beat Ours (193.06/0.2604), and RDGR's ED is also lower. The narrative in §4.4 is more careful — it only claims wins over DocReal on MS-SSIM, LD, and AD — so the bullet needs to be reworded or the numbers rechecked. On DocUNet, the advantage is mainly ED/CER, not the geometric metrics, which also sits awkwardly with an unqualified SOTA sentence.\n\nTwo other issues, both secondary. First, SalmRec is trained on DocDewarpHV1 while most baselines were trained on Doc3D; that is a legitimate design choice for getting clean fine-grained annotations, but it means the benchmark comparisons are not controlled, and part of the gap could be training data rather than architecture. The ablations do not suffer from this — same training set throughout — so the core claim survives. Second, Table 2's baselines are re-evaluated by the authors with their own pipeline, and the OCR engines differ (Tesseract for English, PaddleOCR for Chinese), so a few metric points in either direction are within re-evaluation noise. Code or weights would settle this.\n\nMinor data-quality flags: the Distorted row in Tables 1 and 2 reports identical ED/CER (1500.56/0.5234) on two different benchmarks, one English and one Chinese — that is implausible and looks like a copy-paste error. Table 5's DocUNet ED/CER numbers match the 60-image subset of Table 3 without saying so. Both are easy fixes.\n\nWho is this for: people working on multi-task document dewarping specifically, and anyone using auxiliary tasks to guide geometric prediction. The limitations section is honest (shadows, dense grids, near-background colors) and the citation pattern is fine — the FTDR self-citation is legitimate, since it is their own closest prior work. My recommendation: send it to peer review. The design is worth referee time and the ablation is solid, but the authors should be asked to qualify the SOTA claim, run a controlled comparison on a shared training set (or release code/weights), and fix the table inconsistencies.","headline":"The gating and aggregation modules are credible and the self-ablation holds up, but the unqualified three-benchmark SOTA claim is contradicted by the paper's own DocReal table, and the training-data mismatch keeps the headline result conditional.","tokens_in":17010,"tokens_out":7750,"would_cite":true,"duration_ms":69848,"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":"The paper claims that document rectification reaches state-of-the-art accuracy when four auxiliary tasks—3D coordinates, UV maps, and horizontal/vertical lines—are fused through a leave-one-out feature aggregation module and a gating…","keywords":["document image rectification","dewarping","multitask learning","feature aggregation","gating mechanism","text line segmentation","3D coordinate prediction","UV map"],"falsifier":"Re-run each published baseline checkpoint through the same OCR engine and metric implementation used for Tables 1-3 while retraining that baseline on the same DocDewarpHV1 training set; if the reported margins disappear or reverse, the state-of-the-art claim collapses.","tokens_in":15998,"feed_emoji":"📄","tokens_out":7076,"duration_ms":68173,"temperature":0.7,"pith_summary":"The paper tries to establish that document image rectification improves when auxiliary tasks are fused self-adaptively rather than concatenated. It proposes SalmRec, which predicts 3D coordinates, UV maps, and horizontal/vertical lines, then uses an inter-task feature aggregation module and a gating mechanism to combine them into a 2D deformation field. On DIR300, DocUNet, and DocReal, it reports the best published numbers on several metrics, including the lowest character error rate on DIR300. The significance is that task interference, not lack of auxiliary information, may be the bottleneck.","feed_headline":"Four fused tasks push document dewarping past published baselines","feed_subtitle":"Adaptive fusion of 3D, UV-map, and line features cuts the best reported character error on DIR300 to 14 percent.","key_machinery":"The load-bearing mechanism is the inter-task feature aggregation (FA) module paired with a gating module. FA concatenates three of the four task features, derives a global attention vector from them, multiplies it with the remaining task's feature, and adds a residual connection; this lets each task borrow complementary geometry while suppressing redundant information. The gating module computes channel-wise softmax weights from the shared encoder feature and applies them separately to the global pair (3D/UV) and the local pair (horizontal/vertical lines), before both are concatenated and fed to a transformer decoder that predicts the 2D deformation grid. A weighted line loss supervises the line tasks, and a grid-sampling decoder produces the final rectified image.","core_discovery":"The paper's central claim is that the bottleneck in multi-task document rectification is the way auxiliary tasks are combined, not the auxiliary information itself. SalmRec predicts four complementary signals—3D coordinates, UV texture coordinates, horizontal text lines, and vertical lines—and merges them with an inter-task feature aggregation module that uses a leave-one-out global-attention design, followed by a gating mechanism that weighs global features (3D/UV) and local features (lines) separately. With this combination, the paper reports a state-of-the-art result on the DIR300 benchmark (MS-SSIM 0.67, LD 5.14, AD 0.178, ED 444.07, CER 0.1400), best geometric metrics among compared methods on DocReal, and the best recognition metrics among compared methods in the 60-image OCR setting on DocUNet. The paper reads this as evidence that negative interference between tasks can be reduced by adaptive fusion.","pith_inferences":["The paper trains on DocDewarpHV1 rather than the widely used Doc3D set, so the reported margins mix the effect of the fusion modules with the effect of better line annotations; an ablation that trains the same network and the leading baselines on Doc3D would isolate the modules' true contribution.","Because the failure cases in the paper's own limitation section involve strong shadows, heavy folds, and backgrounds that match the paper edge, the method's practical reach is likely bounded by the same conditions that limit its auxiliary-task predictions.","The leave-one-out aggregation idea is exportable: any set of complementary geometric tasks, such as depth, normal, and edge prediction in generic image restoration, could use the same residual global-attention gating rather than naive concatenation.","A testable compression experiment would replace the 7.1M-parameter gating module with a per-pixel routing layer and check whether the reported gains survive; if they do, the mechanism, not its parameter count, is what matters."],"forward_implications":["Auxiliary supervision is not automatically beneficial: the paper shows that all four tasks together without fusion underperform the fused version on DIR300 and DocReal, so future multi-task dewarpers need explicit interaction modeling.","The reported gains put the best OCR accuracy on DIR300 (CER 0.14) ahead of prior methods that also use text-line attention, suggesting that fused local constraints transfer to recognition.","If the numbers hold, SalmRec becomes the reference point on DIR300 and DocReal, with the fusion and gating modules as the components to ablate in follow-up work.","The architecture's shared encoder and two-gate design imply that additional auxiliary tasks, such as layout masks, could be absorbed by adding one more task branch and one more gate."],"supporting_citations":[{"why":"defines the DIR300 benchmark and the geometric-representation baseline that SalmRec must beat; the paper adopts its evaluation split and comparison.","marker":"[17]"},{"why":"introduces the Doc3D synthetic training set and the 3D-coordinate auxiliary task that anchors the global feature branch.","marker":"[9]"},{"why":"provides the DocUNet benchmark and the stacked-UNet architecture that this work extends.","marker":"[40]"},{"why":"contributes the DocReal Chinese benchmark and the AECP baseline; the paper compares against it and adopts its test protocol.","marker":"[64]"},{"why":"supplies the weighted line loss and grid-regularization idea that the horizontal/vertical line supervision is built on.","marker":"[22]"},{"why":"is the layout/UV-map multi-task baseline against which the global fusion gains are measured.","marker":"[27]"},{"why":"is the text-line cross-attention baseline whose DIR300 CER is improved by 33.81% according to the paper.","marker":"[25]"},{"why":"provides the grid-sampling decoder and transformer unwarping paradigm that SalmRec adopts.","marker":"[15]"},{"why":"provides the global attention operator that the inter-task feature aggregation module applies to the concatenated task features.","marker":"[36]"}],"fun_headline_variants":["SalmRec: adaptive multitask fusion cuts DIR300 CER to 14%","Four-task fusion with gating yields best dewarping on DIR300","Self-adaptive fusion of 3D, UV, and line features tops dewarping baselines","Adaptive gating and aggregation boost document rectification accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline comparison assumes that re-running published baselines through the authors' own evaluation pipeline and training on a different annotated synthetic dataset yields numbers comparable to originally published results.","fun_headline_variants_meta":{"raw":{"variants":["SalmRec: adaptive multitask fusion cuts DIR300 CER to 14%","Four-task fusion with gating yields best dewarping on DIR300","Self-adaptive fusion of 3D, UV, and line features tops dewarping baselines","Adaptive gating and aggregation boost document rectification accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000967,"raw_usage":{"total_tokens":4094,"prompt_tokens":902,"completion_tokens":3192,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":3108}},"tokens_in":518,"tokens_out":3192,"duration_ms":24361,"temperature":1.0,"reasoning_tokens":3108,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:49:06.990270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run each published baseline checkpoint through the same OCR engine and metric implementation used for Tables 1-3 while retraining that baseline on the same DocDewarpHV1 training set; if the reported margins disappear or reverse, the state-of-the-art claim collapses.","supporting_citations":[{"cited_title":"Docunet:Documentimageunwarpingviaastackedu-net","cited_arxiv_id":null,"evidence_quote":"provides the DocUNet benchmark and the stacked-UNet architecture that this work extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"contributes the DocReal Chinese benchmark and the AECP baseline; the paper compares against it and adopts its test protocol."},{"cited_title":"Geometric representation learning for document image rectification, in: European Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"defines the DIR300 benchmark and the geometric-representation baseline that SalmRec must beat; the paper adopts its evaluation split and comparison."},{"cited_title":"Dewarpnet: Single-image document unwarping with stacked 3d and 2d regression networks","cited_arxiv_id":null,"evidence_quote":"introduces the Doc3D synthetic training set and the 3D-coordinate auxiliary task that anchors the global feature branch."},{"cited_title":"Revisitingdocumentimagedewarpingbygridregularization","cited_arxiv_id":null,"evidence_quote":"supplies the weighted line loss and grid-regularization idea that the horizontal/vertical line supervision is built on."},{"cited_title":"Layout-aware single-image document flattening","cited_arxiv_id":null,"evidence_quote":"is the layout/UV-map multi-task baseline against which the global fusion gains are measured."},{"cited_title":"Foreground and text-lines aware document image rectification, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"is the text-line cross-attention baseline whose DIR300 CER is improved by 33.81% according to the paper."},{"cited_title":"Doctr: Document image transformer for geometric unwarping and illumination correction","cited_arxiv_id":null,"evidence_quote":"provides the grid-sampling decoder and transformer unwarping paradigm that SalmRec adopts."}],"review_version":1}