{"id":"767c8951-d7f0-4e0e-8b46-8bb2dbb73cb5","arxiv_id":"1908.11834","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Curved synthetic training data, aspect-ratio-preserving resizing, and rotation augmentation push rectification-based scene text recognizers to state-of-the-art accuracy on curved text benchmarks.","lead":"This paper reports a set of training data and preprocessing tricks that improve scene text recognition on curved and distorted text, including a curved synthetic data generator, aspect-ratio-preserving input resizing, and random rotation. The authors reach state-of-the-art accuracy on curved text benchmarks and won the ICDAR 2019 Arbitrary-Shaped Text recognition challenge (Latin script).","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'combination of techniques' claim is not tested: Table 5 'All' hurts on most datasets, and the ICDAR ensemble is never compared to its best member.","rationale":"In good faith, the paper's core data contribution—CurvedSynth—has real support: within-model ablations in Table 1 show consistent gains on curved datasets, and the authors release code. The domain-gap worry raised by the reader is legitimate but secondary, because the paper evaluates directly on real curved datasets; the absence of a synthetic-to-real distance metric does not by itself invalidate those transfer numbers. The more acute problem is internal to the central claim. The abstract promises a 'bag of tricks' whose combination significantly improves performance, but Table 5's 'All' row is the only direct test of the combined model modifications and it fails on a majority of datasets. The competition result is a black-box ensemble with no member-versus-ensemble comparison, so a reader cannot tell whether the 74.3% win comes from the proposed combination, from one dominant model, or from the larger real-plus-synthetic training set described in Section 5. This concern is repairable by a small controlled experiment, so the verdict remains conditional rather than reject; however, the strong 'combination' wording should not be accepted until the ensemble ablation is supplied.","tokens_in":12571,"tokens_out":11127,"duration_ms":104425,"concrete_test":"On the official IC19-ArT test set, compute the accuracy of each of the four Section 5 models individually from the same checkpoints used in the ensemble, then compute the four-way voting accuracy; if the voting result is not above the best single model, the claim that the combination of techniques produced the 74.3% win is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the headline claim is that the proposed techniques work together to improve irregular text recognition, and the paper's own evidence does not establish this. In Table 5, the row labeled 'All' (Squarization + Random Rotation + Rectify Image) is worse than the unmodified Rectification Baseline on five of nine datasets (IIIT5K 94.2 vs 94.8, SVT 88.1 vs 89.6, IC13 92.0 vs 92.8, Total-Text 76.1 vs 76.3, IC19-ArT 72.0 vs 72.1) and only ties or barely exceeds the baseline on the rest. Section 4.3 itself concedes that 'rectification on images does not seem to be very effective.' The only externally validated strong number, the ICDAR 2019 result in Section 5, comes from a four-model voting ensemble whose individual members are never evaluated on the official test set; therefore there is no evidence that the voting combination is better than its strongest member. If the central claim is read as 'these tricks, taken together, significantly improve irregular text recognition,' Table 5 contradicts it. If it is read more narrowly as 'some tricks help in some configurations,' the competition result does not prove the combination claim. Without a paired comparison of the ensemble against each of its members, the most impressive number in the abstract cannot be attributed to the proposed combination.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a technical report on the ICDAR 2019 Arbitrary-Shaped Text (Latin) recognition track. It presents a set of techniques for rectification-based scene text recognition: (1) CurvedSynth, a modified SynthText engine that generates more curved text; (2) squarization preprocessing plus random rotation; (3) rectification on images; and (4) an ensemble system for the competition. The experiments show that training on CurvedSynth, especially when combined with Synth90K, improves accuracy on curved benchmarks such as CUTE80, Total-Text, and IC19-ArT. The paper also introduces RectTotal, a rectified version of Total-Text. The claimed headline results are 89.6% on CUTE80 and 76.3% on Total-Text, and a 74.3% accuracy in the ICDAR 2019 challenge.","tokens_in":12812,"tokens_out":10267,"duration_ms":83807,"significance":"If the data-side claims hold, CurvedSynth is a useful and economical training-data resource for irregular text recognition. The controlled comparison in Table 1, which holds the architecture fixed and varies only the training set, is a genuine strength and supports the conclusion that curved synthetic data are important. The release of the code and generated data samples is commendable and supports reproducibility. However, the paper's broader claim that the proposed 'bag of tricks' or their combination significantly improves performance is not supported by the ablations in Table 5, where the 'All' configuration is worse than the baseline on most datasets. The ICDAR competition result is obtained by an ensemble, but no ablation isolates the ensemble's contribution. The paper would be a solid contribution if reframed around the CurvedSynth study, but the current abstract and conclusions overstate the evidence for the model modifications and the combination claim.","major_comments":[{"comment":"The 'All' row in Table 5, which combines Squarization, Random Rotation, and Rectify Image, is worse than the unmodified Rectification Baseline on eight of nine datasets (e.g., IIIT5K 94.2 vs. 94.8, SVT 88.1 vs. 89.6, IC13 92.0 vs. 92.8, Total-Text 76.1 vs. 76.3, IC19-ArT 72.0 vs. 72.1), with CUTE80 as the only exception (90.3 vs. 89.6). Section 4.3 itself concedes that 'rectification on images does not seem to be very effective.' This directly contradicts the abstract's claim that the bag of tricks 'significantly improve[s]' performance. The evidence instead shows that the combined model tweaks harm most benchmarks. Please either revise the central claim to state that only some tricks help in some configurations, or provide an ablation of the exact combination used in the competition system.","section":"Section 4.3 / Table 5"},{"comment":"The ICDAR 2019 result (74.3% on the held-out test set) is obtained by a four-model voting ensemble, but the paper reports only the individual validation accuracies (76.5% to 78.5%) and the ensemble's test accuracy. It does not report the ensemble's validation accuracy or the per-model test-set accuracies. Without a paired comparison of the ensemble against each of its members on the same held-out set (or at least on the 3000-image validation set), the 74.3% accuracy cannot be attributed to the proposed combination; it is possible that the strongest single model already achieves this level. The competition result is therefore not evidence for the combination claim.","section":"Section 5"}],"minor_comments":[{"comment":"The squarization preprocessing is not fully specified: after making the input square, the network is still described as taking 64x256 inputs, and the text does not explain how the aspect-ratio-preserving square is reshaped to 64x256 without reintroducing distortion. Please clarify the exact resizing and padding steps, or cite the relevant part of the released code.","section":"Section 4.1"},{"comment":"The paper claims to be 'the first to evaluate on Total-Text' (Section 1) and 'the first to report experiment results' for the latter two curved datasets (Section 2.3), yet Table 2 itself lists CA-FCN's Total-Text accuracy of 61.6% (marked with *). This internal inconsistency should be corrected.","section":"Section 2.3 / Introduction"},{"comment":"There are several typos: 'III5K' should be 'IIIT5K', and 'We would present some very interesting findings' should be 'We present...'.","section":"Section 3.1.2"},{"comment":"In the Recognition Module description, 'Bi-LSTM layers [8] layers' has a duplicated word 'layers'.","section":"Section 2.1"},{"comment":"The synthetic curve parameters are hand-set, and the paper does not provide any distributional analysis of the generated data (e.g., curvature or character-count histograms). Such an analysis would help assess coverage of the long tail of real curved text, though the downstream gains on real benchmarks already provide indirect evidence of transfer.","section":"Section 3.1.1"},{"comment":"The phrase 'validation accuracy from 76.5% and 78.5%' should read 'from 76.5% to 78.5%'.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper reads like a competition technical report rather than a full research article. The CurvedSynth data contribution is supported by a controlled study and the open-sourced code is a clear asset. However, the abstract and conclusion overstate the 'combination of tricks' claim, which is contradicted by Table 5 and not validated for the competition ensemble. A major revision that scopes the claims and adds the missing ensemble ablation is appropriate. The journal should also weigh whether the 'bag of tricks' framing is the most useful packaging for the data-centric contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competition technical report with one solid, reproducible contribution—CurvedSynth—and an abstract that oversells the rest. The data-side claim holds up; the 'bag of tricks' claim does not.\n\nWhat is actually new: the authors modify the SynthText engine to force more curved text (sampling circular trajectories instead of the default parabola), release the generator and 7M crops, and show with the same recognizer and same protocol that training on CurvedSynth beats SynthText and Synth90K on CUTE80, Total-Text, and IC19-ArT. Those gains are large and consistent across both the rectification and non-rectification baselines. That is real evidence, not a fitted curve. The released code and data make it independently checkable, and that should be credited.\n\nThe soft spots are exactly where the reader and stress-test put them. Table 5 shows the 'All' combination (squarization + random rotation + image rectification) is worse than the plain rectification baseline on five of nine benchmarks, and the text itself concedes image rectification 'does not seem to be very effective.' So the abstract's phrase 'bag of techniques that prove to significantly improve performance' is not supported by the ablations. The ICDAR win is an ensemble of four models, and no member is evaluated alone on the official test set, so the 74.3% cannot be attributed to the proposed combination. I also would have liked variance across seeds; every number is a single run, and on CUTE80/Total-Text the differences are sometimes a few points, which matters at these sizes.\n\nNone of this sinks the CurvedSynth contribution. The paper is a technical report, not a polished journal article, and it is written honestly enough to show the weak ablation rather than hide it. If an editor treated the 'combination' claim as load-bearing, the right outcome is revision, not rejection.\n\nBottom line: for anyone doing scene-text recognition, this is worth reading and citing for the data generator. I would accept it for peer review and ask for a rewritten abstract, seed sweeps, and an ensemble-vs-best-member comparison.","headline":"CurvedSynth is a solid, released data contribution that supports the curved-text gains; the 'bag of tricks' framing overreaches Table 5, but the paper deserves a serious referee after revision.","tokens_in":13419,"tokens_out":2400,"would_cite":true,"duration_ms":22665,"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":"The paper shows that rectification-based scene text recognition is limited more by training data than by architecture, and that a modified synthetic dataset with circular curved text lifts accuracy to 89.6% on CUTE-80 and 76.3% on…","keywords":["irregular scene text recognition","curved text","synthetic data generation","rectification","thin-plate spline","squarization","ICDAR 2019 challenge"],"falsifier":"Retrain the same recognizer on a version of CurvedSynth where the circular trajectories are replaced by the original parabolic trajectories while keeping the number of curved crops, fonts, and backgrounds identical, and compare accuracy on Total-Text; if the gap between the two training sets is much smaller than the gain the paper attributes to CurvedSynth, then the circular rendering protocol, not curved data in general, is the active ingredient. Alternatively, vary the sampled radius distribution, for example all tight curves versus all nearly straight, and observe whether Total-Text accuracy responds monotonically to curvature.","tokens_in":12325,"feed_emoji":"🔤","tokens_out":12139,"duration_ms":97385,"temperature":0.7,"pith_summary":"The paper sets out to show that irregular scene text recognition is limited far more by the choice of training data than by the recognition model itself. It takes a standard rectification-based recognizer—one that warps an irregular text image into a horizontal strip before reading it—and shows that existing synthetic datasets, which contain almost no curved text, prevent the rectification module from ever learning to help. By modifying the text-rendering engine to place words on circular rather than parabolic trajectories, the authors create a training set called CurvedSynth, and when it is combined with the standard straight-text corpus, recognition accuracy reaches 89.6% on the curved benchmark CUTE-80 and 76.3% on Total-Text, surpassing the previous best published results by 6.3 and 14.7 percentage points. A sympathetic reader should care because the result suggests that a large part of curved-text recognition performance can be bought with better synthetic data rather than with new network designs.","feed_headline":"Curved training data push scene text accuracy to 89.6 percent","feed_subtitle":"A rectification-based recognizer trained on circular curved text reaches 76.3% on Total-Text and wins ICDAR 2019.","key_machinery":"Two mechanisms carry the argument. The first is a modified text-rendering module for the SynthText engine: instead of the engine's default settings, which almost always sample multi-line straight paragraphs, the module forces single-word rendering and randomly samples a radius so that characters are placed on a circle, producing the CurvedSynth dataset with roughly 10% strongly curved crops among its 7 million images. This data change is what unlocks the second mechanism, the rectification module—a Thin-Plate-Spline warping driven by a predicted 20-vertex control polygon, taken from the Aster architecture—which straightens curved text before an attentional sequence decoder reads it. A third supporting tweak, squarization, resizes the long side of each input to 256 pixels and pads the short side to a square, preserving aspect ratio and making random 90, 180, and 270 degree rotations during training possible; the paper reports that squarization's main value is that it enables this rotation trick.","core_discovery":"The central discovery is that the rectification layer of a scene text recognizer is only as good as the curvature statistics of its training data. When trained on SynthText and Synth90K, whose output is dominated by straight text, the rectification module improves curved-text accuracy by less than a percentage point; when the same model is trained on CurvedSynth—a version of SynthText in which about 10% of the generated crops follow circular trajectories—the rectification layer accounts for improvements of 3.2% on Total-Text and 1.8% on the ICDAR 2019 ArT validation set. The paper further reports that the fully trained model wins the ICDAR 2019 Arbitrary-Shaped Text Challenge (Latin script) with 74.3% accuracy on the held-out test set, and that its RectTotal experiment, obtained by rectifying Total-Text with the TextSnake detector using ground-truth geometry, indicates that a recognizer trained purely on straight text performs equally well once the text has been straightened upstream by detection.","pith_inferences":["The paper leaves open whether the circular-trajectory rendering itself, rather than curved text in general, is what drives the gain; a natural extension would be to generate curved text by sampling from a wider family of smooth curves, such as spirals or arbitrary splines, and measure whether accuracy on real curved images keeps rising.","The optimal 15% real-world data mixing ratio was chosen on a single validation split; one might expect it to depend on the size and domain of the target dataset, so the recipe may not transfer unchanged to benchmarks with very different image statistics.","The RectTotal result suggests a detector-centric route to irregular text recognition: if detectors can output accurate rectified crops, the need for curvature-specialised recognizers may largely disappear, potentially simplifying the whole pipeline even though the paper's own competition system still relies on a curved-aware recognizer.","Because CurvedSynth is generated from the same rendering engine as SynthText, it inherits font, background, and vocabulary biases; testing on real photographs of curved text with unusual fonts or heavy occlusion would clarify how much of the reported gain is curvature per se versus a generally larger and more diverse training corpus."],"forward_implications":["Training on curved synthetic data should replace training on straight-only corpora for every irregular-text benchmark, because the same rectification model improves by 6.3 points on CUTE-80 and 14.7 points on Total-Text without any architectural change.","The rectification layer's value is contingent on training-data curvature, so conclusions about whether rectification helps should be re-examined for models trained on datasets that contain few curved words.","If a detection stage can already output a geometrically rectified crop of the text, a recognizer trained on straight text alone can match more sophisticated curved-text recognizers, shifting design effort toward detection-recognition co-design.","The reported 74.3% held-out accuracy on the ICDAR 2019 challenge shows that the combination of CurvedSynth, squarization, random rotation, and a four-model ensemble transfers from synthetic training to a competition test set."],"supporting_citations":[{"why":"supplies the synthesis engine whose text-rendering module is modified to create CurvedSynth","marker":"[6]"},{"why":"provides the balanced straight-text corpus combined with CurvedSynth in the final training mix","marker":"[9]"},{"why":"defines the rectification-based recognizer architecture used throughout the paper","marker":"[29]"},{"why":"is the large curved-text test set on which the final 76.3% accuracy is reported","marker":"[3]"},{"why":"is the smaller curved-text test set on which the final 89.6% accuracy is reported","marker":"[25]"},{"why":"serves as the strong baseline that also uses real world data, which the paper surpasses","marker":"[14]"},{"why":"provides the prior curved-text state-of-the-art numbers that the method beats","marker":"[16]"},{"why":"supplies the shape-rectification algorithm used to build the RectTotal benchmark","marker":"[21]"},{"why":"is the thin-plate spline transformation the rectification module uses to warp text","marker":"[34]"}],"fun_headline_variants":["Curved training data boosts curved text recognition by 6.3%","Rectification layers need curved samples to improve accuracy","Scene text recognizer wins ICDAR 2019 with curved data trick","Bag of tricks lifts curved text accuracy to 89.6 percent","Training on curved text makes rectification worth it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that text rendered along synthetic circular trajectories with hand-set parameters is a faithful enough proxy for real-world curved text that models trained on CurvedSynth transfer to CUTE-80, Total-Text, and the ICDAR 2019 challenge; the paper does not measure that domain gap.","fun_headline_variants_meta":{"raw":{"variants":["Curved training data boosts curved text recognition by 6.3%","Rectification layers need curved samples to improve accuracy","Scene text recognizer wins ICDAR 2019 with curved data trick","Bag of tricks lifts curved text accuracy to 89.6 percent","Training on curved text makes rectification worth it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1510,"prompt_tokens":948,"completion_tokens":562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":477}},"tokens_in":564,"tokens_out":562,"duration_ms":4898,"temperature":1.0,"reasoning_tokens":477,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:06:02.543562+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same recognizer on a version of CurvedSynth where the circular trajectories are replaced by the original parabolic trajectories while keeping the number of curved crops, fonts, and backgrounds identical, and compare accuracy on Total-Text; if the gap between the two training sets is much smaller than the gain the paper attributes to CurvedSynth, then the circular rendering protocol, not curved data in general, is the active ingredient. Alternatively, vary the sampled radius distribution, for example all tight curves versus all nearly straight, and observe whether Total-Text accuracy responds monotonically to curvature.","supporting_citations":[{"cited_title":"Synthetic data for text localisation in natural images","cited_arxiv_id":null,"evidence_quote":"supplies the synthesis engine whose text-rendering module is modified to create CurvedSynth"},{"cited_title":"Aster: An attentional scene text recognizer with flexible rectification","cited_arxiv_id":null,"evidence_quote":"defines the rectification-based recognizer architecture used throughout the paper"},{"cited_title":"Total-text: A comprehensive dataset for scene text detection and recognition","cited_arxiv_id":null,"evidence_quote":"is the large curved-text test set on which the final 76.3% accuracy is reported"},{"cited_title":"A robust arbitrary text detection system for natural scene images","cited_arxiv_id":null,"evidence_quote":"is the smaller curved-text test set on which the final 89.6% accuracy is reported"},{"cited_title":"Show, attend and read: A simple and strong baseline for irregular text recognition","cited_arxiv_id":null,"evidence_quote":"serves as the strong baseline that also uses real world data, which the paper surpasses"},{"cited_title":"Scene text recognition from two-dimensional perspective","cited_arxiv_id":null,"evidence_quote":"provides the prior curved-text state-of-the-art numbers that the method beats"},{"cited_title":"Textsnake: A flexible representation for detecting text of arbitrary shapes","cited_arxiv_id":null,"evidence_quote":"supplies the shape-rectification algorithm used to build the RectTotal benchmark"},{"cited_title":"Thin-plate splines and the decompositions of deformations","cited_arxiv_id":null,"evidence_quote":"is the thin-plate spline transformation the rectification module uses to warp text"}],"review_version":1}