{"id":"b498f01b-931e-469d-8a62-d47707cbf234","arxiv_id":"1908.09990","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A semi- and weakly-supervised scheme trains curved text detectors with 10% pixel-level labels plus weak rectangle labels, reaching performance near fully supervised state-of-the-art on CTW1500 and Total-Text.","lead":"What if scene text detectors could be trained with mostly boxes instead of precise pixel outlines? This paper shows a curved-text detector can match older state-of-the-art results using only 10% pixel-level labels plus rectangle-level pseudo labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Best-round selection on the test set plus a single random 10% split makes the '10% labels, SOTA-comparable' claim statistically unsupported; a fixed-round/multi-split rerun is needed.","rationale":"The reader's CONDITIONAL verdict already identifies verification weakness, and I agree that the pseudo-mask trust issue in Section III.C3 is a real mechanism-level risk. However, the single most load-bearing problem is not the mask-generation mechanism itself but the evaluation protocol: the headline Local numbers are the maximum F-measure over recursive training rounds selected on the test set, rather than a pre-specified or validation-selected round. With only three rounds and small test sets (500 images for CTW1500, 300 for Total-Text), best-of-three selection can plausibly account for a 2.0-2.5 point gap to the fully supervised upper bound. The single random 10% split adds further uncertainty. A fixed-round, multi-split replication would settle whether the claimed label-efficiency advantage is real or an artifact of test-set selection and split luck. Since the reader's verdict is already CONDITIONAL and this concern reinforces it rather than overturning the paper's contribution, I recommend no change to the verdict.","tokens_in":8352,"tokens_out":4463,"duration_ms":48873,"concrete_test":"Re-run the Local strategy on both datasets under a fixed protocol: choose training round 3 a priori (or choose the round on a held-out validation split, not the test set), repeat the 10%/90% split with at least five random seeds, and report mean and standard deviation of F-measure for Baseline, Naive, Filter, Local, and Fully. If Local's mean F-measure is no longer at least roughly 4 points above Filter and within roughly 2 points of Fully on both datasets, the headline claim is not supported. Also report the round-1 and round-2 F-measures for Local to quantify how much of the reported 76.0/79.4 comes from best-of-three test-set selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the Conclusion is that the method is 'comparable with the state of the art methods with only 10% pixel-level annotated training data and 90% weakly rectangle based annotated training data.' The supporting numbers are Local 76.0 on CTW1500 and 79.4 on Total-Text (Tables I and II). Those numbers are not outputs of a fixed training protocol: Section IV.C states, 'The model with the top performance across all training rounds is picked for evaluation,' and Figures 3-4 show that Local peaks only at the third recursive round. Selecting the maximum of three test-set F-measures makes the evaluation a search over rounds, so the gap to the fully supervised upper bound (78.5/81.4) and to Filter (71.2/76.8) is not a clean estimate of label efficiency. This is fragile because recursive training with noisy pseudo masks could produce round-to-round instability; the selection rule would report the favorable peak. The single 10%/90% split (Section IV.A, one random draw, no seed, no repeats) compounds the issue by giving no variance estimate. The reader's concern about pseudo-mask quality is real, but the evaluation protocol is the more direct threat to the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a semi- and weakly-supervised framework for curved text detection in natural images. A Mask R-CNN-style detector with a coarse localization (bounding box) branch and a fine localization (mask) branch is first trained on a small amount (10%) of pixel-level annotated data. Three strategies then exploit the remaining 90% of images: a naive semi-supervised strategy that thresholds model outputs, a filter strategy that additionally requires high IoU between detected boxes and ground-truth horizontal rectangles, and a local strategy that feeds ground-truth bounding boxes as proposals to the fine localization network to generate pseudo masks. Recursive training is applied. Experiments on CTW1500 and Total-Text show that the local strategy improves F-measure from 66.1 to 76.0 and from 70.1 to 79.4, respectively, approaching the fully supervised upper bounds of 78.5 and 81.4. A cross-dataset experiment transfers a model from CTW1500 to Total-Text and reports 78.1 F-measure. The central claim is that the method is comparable to state of the art with only 10% pixel-level annotations and 90% rectangle-level weak annotations.","tokens_in":8586,"tokens_out":5112,"duration_ms":50063,"significance":"The paper addresses a practical and timely problem: reducing the annotation cost for curved text detection. The local strategy is simple and the cross-dataset transfer experiment is a valuable addition. If the reported results are robust, the contribution would be a useful step toward low-cost arbitrary-shape text detection. However, the evidence as presented is weakened by evaluation protocol choices (best-round selection on the test set, a single random split with no variance estimate) and by a lack of analysis of the pseudo-mask quality on which the local strategy depends. The paper does not release code or machine-checked proofs, but the described experimental setup is reproducible in principle with additional experiments. Because the core idea is plausible and the reported gains are consistent across two datasets, the contribution merits consideration once the evaluation is strengthened.","major_comments":[{"comment":"The paper states that 'the model with the top performance across all training rounds is picked for evaluation.' For the three strategies, the reported F-measures are therefore maxima over the three recursive training rounds on the test set, whereas the fully supervised baseline and the SOTA references are single models. This selection procedure inflates the measured improvement (e.g., 76.0 vs 66.1 on CTW1500) and makes the comparison against the fully supervised upper bound (78.5) not a clean estimate of label efficiency. The authors should report the F-measure at every training round, or at least the mean and standard deviation over rounds, and should compare with fully supervised models evaluated under the same best-round selection rule.","section":"Section IV.C, Tables I and II"},{"comment":"For both datasets, a single random 10%/90% split is used (100/900 for CTW1500 and 125/1130 for Total-Text), with no repeated splits or random seeds. Consequently, there is no variance estimate for the headline claim that 10% pixel-level labels plus 90% weak rectangle labels suffice to approach SOTA. The differences between the local strategy (76.0/79.4) and the fully supervised upper bound (78.5/81.4) may be within the noise of the particular split. The authors should repeat the experiments with multiple random splits and report mean and standard deviation.","section":"Section IV.A, Tables I and II"},{"comment":"The local strategy treats as ground truth the masks produced by the fine localization network from ground-truth rectangles, but this network is trained on only 10% of the images. No analysis is provided of the quality of these pseudo masks (e.g., mask IoU with human annotations on a held-out subset), nor of how masking errors propagate through the recursive training described in Section III-D. Since the reported gains depend directly on the pseudo masks being reliable enough to train from, the authors should include such a quality analysis or an oracle experiment that replaces pseudo masks with ground-truth masks to establish an upper bound.","section":"Section III-C3, Eq. (4), and Section III-D"}],"minor_comments":[{"comment":"The thresholds S (0.5), S' (0.4), and T (0.3) are set empirically without any sensitivity analysis; because they directly control the pseudo-label sets for the naive and filter strategies, a small ablation study would strengthen the paper.","section":"Section IV.B"},{"comment":"The cross-dataset experiment does not state whether the best training round is selected on the Total-Text test set; if so, the same best-round selection issue applies and should be clarified.","section":"Section IV.C"},{"comment":"The prediction target for the fine localization network is described only as 'the intersection between an ROI and its associated ground-truth mask'; an explicit equation or figure would improve reproducibility.","section":"Section III.B"},{"comment":"Reference [3] contains an author-name typo: 'C. Ya' should be 'C. Yao'.","section":"References"},{"comment":"The figures show F-measure versus training round for the three strategies, but the line styles are not defined in a legend, and the y-axis range is not specified; adding a legend and axis range would improve readability.","section":"Figures 3 and 4"},{"comment":"The claim of being 'comparable with state-of-the-art' should be qualified because the cited SOTA methods use different backbones and training data; the fully supervised rows of the same framework provide a fairer reference point.","section":"Tables I and II"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible annotation-efficient method, but the current evidence does not yet support the headline statistical claim. The main concerns are the best-round test-set selection and the single random split. I recommend asking the authors to rerun with multiple splits and to report per-round performance, and to add a pseudo-mask quality analysis. These are addressable within the scope of the paper, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline claim—comparable with state of the art using only 10% pixel-level annotated data—is plausible but not actually established by the experiments as reported. The local strategy (feed ground-truth boxes into the mask head to synthesize pseudo masks) is a sensible twist on self-training, and the cross-domain CTW1500→Total-Text result is a nice practical bonus. The paper deserves a referee, but the evaluation protocol needs tightening before the central claim can be trusted.\n\nWhat's genuinely new: the local strategy is not in WeText, Wordsup, or SEE. Instead of thresholding the model's own box/mask outputs, it uses the weak rectangle annotations as proposals and lets the fine localization network produce the mask. That's a reasonable division of labor: the coarse box is trusted, the mask is inferred. The 76.0/79.4 F-measures against fully supervised 78.5/81.4 in the same framework are credible evidence that label efficiency exists. The authors also correctly credit WeText as the closest work and position their contribution as extending it to arbitrary/curved text.\n\nThe soft spots are concentrated in the evaluation, and they are load-bearing. The model 'with the top performance across all training rounds is picked for evaluation'—that is test-set selection over the recursive round. Figures 3–4 show local peaks only at round three, so the reported number is the maximum of a small search, not the output of a fixed protocol. Combined with a single random 10%/90% split and no error bars, the gap between Local and Filter/Naive is not clean evidence for the method. The thresholds S, S′, and T are set empirically, with no ablation; if the results are sensitive to them, the robustness story weakens further. The mask-quality concern is real but secondary: if the mask head produces biased masks, recursive training can amplify the bias, and best-round selection would hide exactly that instability. No code is released, which makes re-running the protocol impossible.\n\nThe citation pattern is fine; no sign of missing prior work or stacked self-citations. The framework section is clearly written and the figures match the tables.\n\nWho is this for? Anyone working on label-efficient scene text detection, and more broadly on self-training for instance segmentation. As a claim that 10% pixel labels are enough, it needs a multi-split, fixed-round re-run. I would not desk-reject it; I would send it to review and ask for that experiment.","headline":"A sensible pseudo-mask strategy for label-efficient curved text detection, but the headline claim is built on best-round test selection and a single split; worth reviewing with a request for a fixed-protocol rerun.","tokens_in":9124,"tokens_out":2250,"would_cite":false,"duration_ms":21397,"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 a curved text detector can be trained to near state-of-the-art accuracy with only 10% pixel-level annotated images and 90% rectangle-annotated images, by generating pseudo masks from ground-truth boxes.","keywords":["curved text detection","weakly supervised learning","semi-supervised learning","pseudo mask annotation","scene text detection","instance segmentation","rectangle annotations","recursive training"],"falsifier":"Take the weakly annotated images used in the local strategy, compare each generated pseudo mask against withheld polygon ground truth by IoU, and check whether low agreement correlates with the final F-measure gap; if the pseudo masks match the polygons closely, the mechanism is confirmed, and if they are consistently off (missing text tails, including background), the claimed savings would not survive in other settings.","tokens_in":8137,"feed_emoji":"🔤","tokens_out":7493,"duration_ms":70476,"temperature":0.7,"pith_summary":"The paper asks whether accurate curved-text detection, which usually depends on expensive pixel-level polygon annotations, can be trained mostly from cheap rectangle annotations. It proposes a semi- and weakly-supervised framework in which a baseline detector trained on 10% pixel-level annotated images annotates the remaining 90% of weakly annotated or unlabeled images, and a 'local' strategy feeds ground-truth bounding boxes into the mask branch to produce pseudo masks. On CTW1500 and Total-Text, the local strategy reaches 76.0 and 79.4 F-measure, close to the fully supervised upper bound of 78.5 and 81.4 in the same framework and comparable to state-of-the-art methods trained with full annotations. The claim matters because it suggests segmentation-based curved text detectors can be deployed without large pixel-level labeling efforts.","feed_headline":"Curved text detection needs only 10% pixel-level labels","feed_subtitle":"A detector trained on cheap rectangle boxes reaches near fully supervised accuracy on curved text benchmarks.","key_machinery":"The load-bearing mechanism is pseudo-mask generation from ground-truth bounding boxes. The fine localization network learns a mapping from region features sampled inside a box to the binary mask of the text instance; when the weakly annotated images supply only horizontal rectangles, those rectangles are fed directly into this network as proposals, and the network's mask output, paired with the ground-truth box, becomes the pseudo annotation used to retrain the detector. In the local strategy, every ground-truth box $g_i$ yields a pseudo mask $m_i$, and $P=\\{(g_i,m_i)\\mid g_i\\in G\\}$ is the pseudo annotation set. The box provides localization while the mask branch supplies the curved shape, converting cheap rectangle supervision into dense mask supervision.","core_discovery":"On the paper's own terms, the central discovery is that the expensive pixel-level annotation of curved text can be largely replaced by rectangle annotations without losing much accuracy. Using a detector with a coarse localization branch that predicts horizontal boxes and a fine localization branch that predicts text instance masks, the authors train a baseline on 10% pixel-level annotated data, then treat each ground-truth rectangle in the weakly labeled data as a proposal for the fine localization branch and use the predicted mask as a pseudo annotation. Recursive training with these pseudo masks lets the local strategy reach 76.0 F-measure on CTW1500 and 79.4 on Total-Text, versus 78.5 and 81.4 for full supervision, and a 78.1 F-measure in a cross-dataset setting where the model trained on CTW1500 annotates Total-Text. The paper interprets the local strategy's slower saturation across training rounds as evidence that its pseudo masks are cleaner than those produced by naive or filter-based strategies.","pith_inferences":["If the pseudo masks stay accurate when boxes are not tight horizontal rectangles, the same trick could extend to quadrilateral or rotated-box annotations, which would cover more weakly supervised benchmarks; the paper only tests horizontal boxes.","The 10% pixel-level fraction is likely not a magic threshold; a natural extension is to measure how performance degrades as that fraction shrinks toward 1%, which would map the annotation savings more precisely.","The cross-dataset result suggests a practical workflow: keep a pixel-level annotated source set, run the local strategy on any rectangle-annotated target set, and adapt the detector to a new text domain without drawing new polygons.","Directly measuring pseudo-mask quality (for example, IoU against withheld polygon labels) would isolate whether residual errors come from mask generation or from later recursive amplification, a distinction the paper does not test."],"forward_implications":["With 10% pixel-level annotated data and 90% rectangle-annotated data, the local strategy reaches 76.0 F-measure on CTW1500 and 79.4 on Total-Text, within about 2 points of fully supervised training in the same framework.","The local strategy outperforms the naive and filter strategies by 4.8 and 2.6 F-measure points on CTW1500, respectively, and keeps improving through more training rounds, indicating its pseudo masks carry less noise.","All three strategies beat the baseline trained only on the small pixel-level set, so the framework's recursive self-annotation provides a consistent boost.","In a cross-dataset experiment, the local strategy transfers a detector from CTW1500 to Total-Text using only rectangle annotations on the target set, reaching 78.1 F-measure, close to a leading fully supervised curved-text detector on that benchmark."],"supporting_citations":[{"why":"Supplies the curved text detector architecture (feature extractor, proposal network, region sampler, coarse and fine localization) that the framework extends.","marker":"[14]"},{"why":"Provides the region-feature sampling and mask-branch design that the detector's fine localization network is built on.","marker":"[16]"},{"why":"Contributes the weakly supervised filtering idea that the filter strategy adapts for rectangle-annotated text data.","marker":"[13]"},{"why":"Serves as a state-of-the-art curved text detector reference whose reported F-measures anchor the comparison on both datasets.","marker":"[8]"},{"why":"Serves as the state-of-the-art reference for curved text detection on Total-Text.","marker":"[9]"},{"why":"Introduces the CTW1500 benchmark and a curved text detection baseline used for evaluation and comparison.","marker":"[7]"},{"why":"Introduces the Total-Text benchmark used to evaluate the method's generalization to curved, multi-oriented, and horizontal text.","marker":"[6]"}],"fun_headline_variants":["10% pixel labels match full supervision for curved text","Rectangle boxes replace pixel masks in curved text detection","Weak supervision cuts curved text annotation effort by 90%","Pseudo masks from boxes train curved text detector","Curved text detection with minimal pixel-level labeling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the mask branch, trained on just 10% pixel-level data, generalizing to the weakly annotated images well enough that the pseudo masks it produces from ground-truth boxes are accurate rather than systematically wrong.","fun_headline_variants_meta":{"raw":{"variants":["10% pixel labels match full supervision for curved text","Rectangle boxes replace pixel masks in curved text detection","Weak supervision cuts curved text annotation effort by 90%","Pseudo masks from boxes train curved text detector","Curved text detection with minimal pixel-level labeling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1499,"prompt_tokens":901,"completion_tokens":598,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":524}},"tokens_in":517,"tokens_out":598,"duration_ms":6102,"temperature":1.0,"reasoning_tokens":524,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:57:14.116952+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the weakly annotated images used in the local strategy, compare each generated pseudo mask against withheld polygon ground truth by IoU, and check whether low agreement correlates with the final F-measure gap; if the pseudo masks match the polygons closely, the mechanism is confirmed, and if they are consistently off (missing text tails, including background), the claimed savings would not survive in other settings.","supporting_citations":[{"cited_title":"Fused text segmentation networks for multi-oriented scene text detection,","cited_arxiv_id":null,"evidence_quote":"Serves as the state-of-the-art reference for curved text detection on Total-Text."},{"cited_title":"Towards end-to-end text spotting with convolu- tional recurrent neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the curved text detector architecture (feature extractor, proposal network, region sampler, coarse and fine localization) that the framework extends."},{"cited_title":"Mask r-cnn,","cited_arxiv_id":null,"evidence_quote":"Provides the region-feature sampling and mask-branch design that the detector's fine localization network is built on."},{"cited_title":"Wetext: Scene text detection under weak supervision,","cited_arxiv_id":null,"evidence_quote":"Contributes the weakly supervised filtering idea that the filter strategy adapts for rectangle-annotated text data."},{"cited_title":"Textsnake: A flexible representation for detecting text of arbitrary shapes,","cited_arxiv_id":null,"evidence_quote":"Serves as a state-of-the-art curved text detector reference whose reported F-measures anchor the comparison on both datasets."},{"cited_title":"Total-text: A comprehensive dataset for scene text detection and recognition,","cited_arxiv_id":null,"evidence_quote":"Introduces the Total-Text benchmark used to evaluate the method's generalization to curved, multi-oriented, and horizontal text."}],"review_version":1}