{"id":"f3057b24-7766-45d3-ab51-9d29a22ef7e0","arxiv_id":"2504.20682","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A YOLO-based instance segmentation model with gradient-orientation features and mask-based suppression reaches the best cell segmentation accuracy on the new synthetic DWTAL deformed-table datasets.","lead":"This paper builds a deep-learning model that finds and outlines each cell in photos of bent or folded tables, and releases a large synthetic training dataset. The value is that downstream software can extract the exact content of each cell even when the table is warped by camera angle, folds, or shadows.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DWTAL random image-level split may leak source tables into both train and test, inflating the reported mAP and the margin over baselines.","rationale":"The reader's weakest assumption is exactly the load-bearing issue. The central claim is not merely that the proposed modules work in isolation but that OG-HFYOLO is superior for deformed-table cell segmentation, operationalized on DWTAL. If the test set contains warped copies of tables whose other copies were seen in training, the evaluation measures layout memory rather than deformation-robust generalization. In the limit where each source table contributes many variants, any high-capacity model can appear to generalize. The parameter-count asymmetry between OG-HFYOLO (125.39M) and YOLOv5l-seg (92.27M) strengthens the concern, since a larger model has more capacity to exploit source-table identity; therefore the relative gains, not just absolute numbers, are at risk. I agree with the reader's assessment. The contribution is checkable because the dataset and code are open source, so the appropriate posture is CONDITIONAL rather than REJECT: there is an unresolved risk, not demonstrated failure. I would not accept unconditionally until the source-grouped split is run, and I would not reject because the paper is transparent enough to allow the test. Minor issues such as the Eq. 18 numerator typo and the unspecified mask-NMS threshold are worth fixing but do not alter the verdict.","tokens_in":17802,"tokens_out":5659,"duration_ms":57637,"concrete_test":"Download the released DWTAL annotations and generator script from https://github.com/justliulong/OGHFYOLO and recover a source-table identifier for every generated image, either from filenames, metadata, or a generation log. Rebuild the train/test split so that all generated variants of the same source table are assigned to the same fold. Then retrain OG-HFYOLO and the strongest baseline YOLOv5l-seg from scratch using the paper's hyperparameters (SGD, 200 epochs, 640x640 input, batch size 2 for DWTAL-s, batch size 1 for DWTAL-l) and compare Mask mAP@50:95 on DWTAL-s and DWTAL-l with Tables 2 and 3. If the proposed model's margin narrows materially or the absolute mAP drops, the reported superiority is partly a leakage artifact. If source identifiers are not recoverable, run the generator with logging to count how many images each source table produces and repeat the grouped split.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is the train/test split of the DWTAL datasets, described in Section 3.2 as an image-level random 80/20 allocation. Section 3.1 and Section 3.2 explain that mildly deformed tables from TAL-OCR/WTW are converted to segmentation masks and then 'iteratively processed by the data generator to produce diverse deformed table images.' Consequently, multiple generated images can share the same original table layout. A random image-level split can place differently warped versions of the same source table in both the training and test sets, allowing the model to memorize table layouts rather than learn deformation-robust cell boundaries. This is especially concerning because the proposed model has 125.39M parameters versus 92.27M for YOLOv5l-seg, giving it more capacity to exploit source-table identity; both the absolute mAP values and the 2.27-point DWTAL-s margin over YOLOv5l-seg could be inflated. The paper provides no grouping by source table and no statement that source tables are disjoint across splits. Since the data and code are open source, this is directly checkable. Secondary issues, such as the Eq. 18 numerator typo and the unspecified mask-NMS threshold, are real but less central.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes OG-HFYOLO, a YOLO-based instance segmentation model for deformed table cells, combining a Gradient Orientation-aware Extractor (GOE), a Heterogeneous Kernel Cross Fusion (HKCF) module, a scale-aware loss, and mask-driven non-maximal suppression. It also introduces a data generator that produces deformed table images from TAL-OCR, WTW, and 150 offline images, yielding the DWTAL-s and DWTAL-l datasets. Experiments report mask mAP@50:95 of 74.23% on DWTAL-s and 62.38% on DWTAL-l, exceeding the best cited baseline, YOLOv5l-seg, by 2.27 and 1.04 points respectively. The authors claim state-of-the-art segmentation accuracy among mainstream instance segmentation models on the DWTAL dataset.","tokens_in":18089,"tokens_out":5002,"duration_ms":48146,"significance":"If the reported results are trustworthy, the work provides a useful benchmark for fine-grained deformed table cell localization and a plausible architectural recipe for dense, scale-variant instance segmentation. The open-sourcing of the dataset and code is a clear strength. However, the evaluation protocol has a serious potential flaw: because the DWTAL images are generated from a limited set of source tables, a random image-level split can place differently warped versions of the same source table into both training and test sets. This would inflate the reported mAP values and the margins over baselines. The comparison with baselines is also not controlled for training budget. The central claim therefore requires verification under a corrected split before the numerical results can be accepted.","major_comments":[{"comment":"Section 3.2 states that DWTAL-s and DWTAL-l are split by randomly allocating 80% of images to training and 20% to testing. Since Section 3.1 describes generating many deformed images from each selected source table (TAL-OCR, WTW, and 150 offline tables), a random image-level split can place different warped versions of the same source table in both the training and test partitions. This permits the model to memorize table layouts rather than learn deformation-robust cell boundary segmentation. As a result, the absolute mAP numbers in Tables 2 and 3, and especially the 2.27-point margin over YOLOv5l-seg on DWTAL-s, may be substantially inflated. The authors should either split at the source-table level and re-run all experiments, provide explicit evidence that no source table contributes images to both partitions, or report results on held-out real tables from different sources.","section":"3.2"},{"comment":"The comparison is not controlled for training schedule or initialization. Section 5.1 states that the proposed model is trained for 200 epochs from scratch, while the non-YOLO baselines (Mask R-CNN, Cascade Mask R-CNN, Mask2Former, SOLOv2, YOLACT) are fine-tuned for 100 epochs from ImageNet-pretrained ResNet-101. The training schedule for YOLOv5l-seg, YOLOv8l-seg, and YOLOv11l-seg is not reported. The gains shown in Tables 2 and 3 could therefore be partially attributable to training budget disparity rather than architectural merit. The authors should train all methods under the same schedule and initialization conditions, or at minimum report the exact epochs and pretraining settings for every baseline and discuss the potential influence on the comparison.","section":"5.1"},{"comment":"All reported performance numbers appear to come from a single training run without standard deviations or repeated-seed experiments. This is particularly concerning for the DWTAL-l comparison, where the margin over YOLOv5l-seg in mask mAP@50:95 is only 1.04 points. Without variance estimates, it is impossible to assess whether the observed improvements are statistically meaningful. The authors should run at least three seeds (or report confidence intervals) for the main comparisons in Tables 2 and 3.","section":"5.3"}],"minor_comments":[{"comment":"Equation (18) defines Mask_IoU as |Mi∩Mi|/|Mi∪Mj|, but the numerator should be |Mi∩Mj|. The notation also switches between Mi and Mi, which should be fixed to distinguish the two masks clearly.","section":"4.5"},{"comment":"The mask-NMS post-processing is described as suppressing a low-confidence mask only if its Mask_IoU with a higher-confidence mask 'exceeds the threshold', but the threshold value is never specified, nor is its sensitivity analyzed. This is a reproducibility gap that should be filled.","section":"4.5"},{"comment":"The number of orientation bins n used in Equations (9) and (10) for the orientation attention mechanism is not reported. Please provide the value and any ablation of this hyperparameter.","section":"4.2"},{"comment":"Table 5 is difficult to read because the checkmark symbols for the ablated components run together in the extracted text. Please reformat the table so each configuration row clearly shows which modules are enabled or disabled.","section":"5.5.2"},{"comment":"The paper introduces the weighting W_s = 1+log(1/A) as a smoother alternative to 1/A, but Equation (17) uses the product (1+log(1/A_i))·(1/A_i) rather than substituting W_s for 1/A. The relationship between the proposed weight and Equation (17) should be clarified, and the theoretical claims about the derivative behavior should be reconciled with the actual loss formula.","section":"4.4"},{"comment":"There is a duplicated sentence in the introductory paragraph of Section 3.1: 'Specifically, the generator incorporates two distortion techniques: wave warping and cylindrical warping.' appears twice. Please remove the duplicate.","section":"3.1"}],"recommendation":"major_revision","confidential_remarks":"The source-table leakage concern in Section 3.2 is the most serious issue and is directly checkable from the released code and dataset. If an audit confirms that images from the same source table appear on both sides of the split, the reported numbers and the paper's central claim would be invalid, and a reject would be warranted. I recommend asking the authors to provide source-level disjointness verification or a corrected split before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWorth reading, but don't trust the numbers yet. The genuine contribution is the DWTAL dataset and generator: pixel-level cell masks for deformed wired tables, plus an open-source synthetic pipeline that fills a real annotation gap. That alone justifies a serious look. The GOE, HKCF, scale-aware loss, and mask-NMS are sensible engineering adaptations of known pieces (HOG, YOLO-MS, EIoU, SOLOv2), not fundamentally new ideas, but they're applied to a new task with open code, which makes the work checkable and useful.\n\nThe soft spot is exactly where the stress-test note lands. Section 3.2 says the split is a random 80/20 allocation over images, after the generator iteratively produces multiple deformed versions of the same source table. If images from one source table land in both train and test, the model can memorize layout rather than learn deformation-robust boundaries. That would inflate the 74.23% mAP and the 2.27-point margin over YOLOv5l-seg. The paper doesn't state that source tables are disjoint. The code is public, so this is directly verifiable — I'd want that check before believing the headline. This is a load-bearing weakness, but it's fixable and not evidence of bad faith.\n\nSecondary issues are real but minor: Equation 18 has a typo (M_i over both numerator and denominator), the mask-NMS IoU threshold is unspecified, and the YOLOv8/YOLOv11 baselines are trained with schedules that differ from the proposed model, making those comparisons apples-to-oranges. The YOLOv5 comparison is fairer since same pipeline was used. Also, the Table 5 ablation table formatting is mangled in this version, which hurts readability but doesn't hide the trend.\n\nWho this is for: anyone working on table structure recognition or document parsing who needs a pixel-level deformed-cell benchmark and a working baseline. If the split is clean, the paper is a solid contribution to that community; if not, it's still a useful dataset paper with inflated accuracy claims.\n\nMy recommendation: engage with it. Send it to peer review, but require the authors to confirm source-table disjointness (or re-split), report the mask-NMS threshold, and fix Equation 18. The core engineering is credible and the dataset is valuable; the evaluation just needs to be made honest.","headline":"A real engineering contribution wrapped in an evaluation that needs a source-leakage check before any of the headline mAP numbers can be taken at face value.","tokens_in":18580,"tokens_out":583,"would_cite":true,"duration_ms":7778,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A YOLO-style cell segmenter claims top accuracy on deformed tables via gradient orientation, cross-scale fusion, scale-aware loss, and mask-NMS.","keywords":["table structure recognition","instance segmentation","deformed table cell localization","gradient orientation","heterogeneous kernel fusion","scale-aware loss","mask non-maximum suppression","data generator"],"falsifier":"Re-split DWTAL by source-table identity so all warped variants of one original table stay in a single partition, retrain OG-HFYOLO and the leading baselines, and compare Mask mAP@50:95; if the margin over YOLOv5l-seg collapses or the absolute mAP drops by more than a few points, the deformation-robust generalization claim is not supported.","tokens_in":17602,"feed_emoji":"📊","tokens_out":6670,"duration_ms":59724,"temperature":0.7,"pith_summary":"The paper is trying to establish that pixel-level instance segmentation can reliably locate cells in geometrically deformed tables, where traditional corner-point detection loses content and wire-table text-box segmentation fails. It proposes OG-HFYOLO, a YOLO-style anchor-based model that adds a gradient-orientation extractor, a heterogeneous-kernel cross-fusion module, a scale-aware loss, and mask-driven non-maximum suppression, and reports that it beats all compared mainstream instance segmentation models on the derived DWTAL datasets. A sympathetic reading takes away two main claims: the model's reported Mask mAP@50:95 of 74.23% on DWTAL-s and 62.38% on DWTAL-l, and the usefulness of a deformable-table data generator that turns coarse corner annotations into pixel-level masks at scale. If these claims hold, downstream table-content extraction gains a finer-grained spatial coordinate source than contour points, and the field gains a benchmark dataset with the missing mask-level annotations.","feed_headline":"Cell segmentation on deformed tables beats mainstream models","feed_subtitle":"Gradient orientation, cross-scale fusion, and mask-NMS lift Mask mAP to 74.23% on DWTAL-s.","key_machinery":"The load-bearing mechanism is a YOLO-style anchor-based detector retrofitted with four interacting components, plus a data generator. The Gradient Orientation-aware Extractor (GOE) decouples horizontal and vertical gradient responses, computes gradient magnitude and an orientation-channel encoding, and applies an orientation attention with kernels initialized from directional basis vectors, so the shallow layers see edge direction and strength. The Heterogeneous Kernel Cross Fusion (HKCF) module uses a bottleneck with channel attention and parallel horizontal/vertical cross-convolutions of kernel sizes 3, 5, and 7 to capture wide-spanning and tall-narrow merged cells. The scale-aware loss replaces the raw inverse-area normalization with $(1+\\log(1/A))$ weighting to smooth gradient growth for small cells. Mask-driven NMS suppresses a detection only when its predicted-mask IoU with a higher-confidence mask exceeds a threshold, avoiding box-overlap artifacts in dense layouts. The DWTAL data generator synthesizes training images through wave warping, cylindrical warping, and corner-weighted illumination adjustment, converting coarse corner annotations into pixel-level masks.","core_discovery":"On its own terms, the paper's discovery is that the hard part of deformed-table cell localization is not finding cells but drawing their exact boundaries, and that a single-stage anchor-based segmenter can be equipped to do this. The proposed OG-HFYOLO couples a Gradient Orientation-aware Extractor (GOE) in the shallow backbone, a Heterogeneous Kernel Cross Fusion (HKCF) module in the neck, a log-scaled scale-aware mask loss, and mask-IoU-based NMS in place of box NMS. On DWTAL-s it reports Mask mAP@50:95 of 74.23%, ahead of the best compared baseline YOLOv5l-seg at 71.96%; on DWTAL-l it reports 62.38%, ahead of YOLOv5l-seg at 61.34%. The paper also introduces the DWTAL dataset, generated by wave and cylindrical warping plus illumination adjustment from mildly deformed tables in WTW, TAL-OCR, and 150 collected images, to provide the first large-scale fine-grained cell-mask benchmark for deformed tables.","pith_inferences":["If the random image-level split lets differently warped versions of the same source table appear in both training and test sets, the reported mAP margins may overstate true generalization to unseen table layouts; a table-level split would settle this directly.","Mask-driven NMS is a transferable idea: any dense-instance domain where bounding boxes overlap heavily, such as medical cell clusters or document layout objects, could be evaluated with the same mask-IoU suppression rule.","The $(1+\\log(1/A))$ loss weighting is a general replacement for the standard inverse-area normalization and could be tested on other instance segmentation benchmarks with extreme scale ratios.","The data generator itself, with its parameter-coupling rules for wave and cylindrical warping and illumination gradients, could serve as a deformation-augmentation recipe for other document analysis tasks such as text detection or layout parsing."],"forward_implications":["Cell localization in deformed tables can move from four contour corner points to pixel masks, giving content extraction a way to keep text that falls near strongly curved boundaries.","The DWTAL dataset supplies the first large-scale mask-level benchmark for deformed wired tables, so subsequent models can be trained and compared on fine-grained spatial coordinates rather than coarse boxes.","Anchor-based detection is measurably better than anchor-free for this task: the paper's ablation shows over 10% mAP improvements from the anchor-based formulation, so future deformed-table detectors should not default to anchor-free.","The proposed components are only effective together: GOE, HKCF, scale-aware loss, and mask-NMS each give small isolated gains, but their combination produces the reported jump, indicating the dense-object and scale-variation problems are coupled."],"supporting_citations":[{"why":"Supplies the WTW deformed-table benchmark whose coarse corner annotations are refined into segmentation masks for the DWTAL-l subset.","marker":"[7]"},{"why":"Supplies the TAL-OCR dataset used as the main source for the simpler DWTAL-s subset.","marker":"[8]"},{"why":"YOLOv5l-seg is both the base architecture OG-HFYOLO builds on and the strongest reported baseline it must beat.","marker":"[20]"},{"why":"YOLO-MS contributes the Heterogeneous Kernel Selection Protocol that the HKCF module adapts for multi-scale feature fusion.","marker":"[10]"},{"why":"SOLOv2's mask competition strategy inspires the mask-driven non-maximum suppression that replaces box IoU suppression.","marker":"[43]"},{"why":"EIoU loss replaces CIoU for bounding-box regression, decoupling width and height optimization for extreme aspect ratios.","marker":"[40]"},{"why":"Dice loss is combined with binary cross-entropy in the base mask loss to emphasize region overlap for cell masks.","marker":"[41]"},{"why":"HOG's gradient orientation and magnitude binning supplies the core principle behind the Gradient Orientation-aware Extractor.","marker":"[33]"}],"fun_headline_variants":["Gradient-guided masks out-segment YOLOv5 on deformed tables","New benchmark DWTAL pushes deformed-table cell masks to 74% mAP","OG-HFYOLO: orientation cues and mask-NMS sharpen cell boundaries","Single-stage segmenter with mask-IoU NMS edges out YOLOv5-seg","DWTAL dataset: first fine-grained deformed-table cell-mask benchmark"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains depend on the test set containing genuinely new table layouts; because the DWTAL split randomly assigns generated images without tracking which source table each deformed image came from, the model could be learning to recognize familiar layouts rather than learning to segment any deformed table.","fun_headline_variants_meta":{"raw":{"variants":["Gradient-guided masks out-segment YOLOv5 on deformed tables","New benchmark DWTAL pushes deformed-table cell masks to 74% mAP","OG-HFYOLO: orientation cues and mask-NMS sharpen cell boundaries","Single-stage segmenter with mask-IoU NMS edges out YOLOv5-seg","DWTAL dataset: first fine-grained deformed-table cell-mask benchmark"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000645,"raw_usage":{"total_tokens":2969,"prompt_tokens":951,"completion_tokens":2018,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":1914}},"tokens_in":567,"tokens_out":2018,"duration_ms":14295,"temperature":1.0,"reasoning_tokens":1914,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:23:16.413034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-split DWTAL by source-table identity so all warped variants of one original table stay in a single partition, retrain OG-HFYOLO and the leading baselines, and compare Mask mAP@50:95; if the margin over YOLOv5l-seg collapses or the absolute mAP drops by more than a few points, the deformation-robust generalization claim is not supported.","supporting_citations":[{"cited_title":"Parsing table structures in the wild","cited_arxiv_id":null,"evidence_quote":"Supplies the WTW deformed-table benchmark whose coarse corner annotations are refined into segmentation masks for the DWTAL-l subset."},{"cited_title":"Tal-ocr-table dataset","cited_arxiv_id":null,"evidence_quote":"Supplies the TAL-OCR dataset used as the main source for the simpler DWTAL-s subset."},{"cited_title":"Jocher et al","cited_arxiv_id":null,"evidence_quote":"YOLOv5l-seg is both the base architecture OG-HFYOLO builds on and the strongest reported baseline it must beat."},{"cited_title":"Yolo-ms: Rethinking multi-scale representation learning for real-time object detection","cited_arxiv_id":null,"evidence_quote":"YOLO-MS contributes the Heterogeneous Kernel Selection Protocol that the HKCF module adapts for multi-scale feature fusion."},{"cited_title":"Focal and efficient iou loss for accurate bounding box regression","cited_arxiv_id":null,"evidence_quote":"EIoU loss replaces CIoU for bounding-box regression, decoupling width and height optimization for extreme aspect ratios."},{"cited_title":"V-net: Fully convolutional neural networks for volumetric medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Dice loss is combined with binary cross-entropy in the base mask loss to emphasize region overlap for cell masks."},{"cited_title":"Dalal and B","cited_arxiv_id":null,"evidence_quote":"HOG's gradient orientation and magnitude binning supplies the core principle behind the Gradient Orientation-aware Extractor."}],"review_version":1}