{"id":"8873a775-7d79-46ef-85f3-1c46a2cdd684","arxiv_id":"2507.20188","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SAViL-Det combines CLIP, an asymptotic feature pyramid, and cross-modal attention to report F-scores of 84.8 on MLT-2019 and 90.2 on CTW1500, claiming state-of-the-art multi-script and curved text detection.","lead":"This paper presents SAViL-Det, a system that uses a vision-language model to find text in photos across many languages and curved shapes. It reports slightly better accuracy than prior methods on two text-detection benchmarks by combining language clues with visual features.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CTW1500 'No textual features' ablation in Table 5 exactly duplicates TextSnake's published R/P/F, so the evidence that textual features are essential is not credible.","rationale":"I read the paper as making two nested claims: (1) SAViL-Det achieves state-of-the-art F-scores on MLT-2019 and CTW1500; (2) the improvement is caused by the language-vision decoder and text-to-pixel contrastive learning, i.e., by injecting textual semantic features. Claim (2) is supported mainly by the 'Impact of text prompt features' ablation in Sec. 4.5, Table 5. For claim (2) to hold, the 'No textual features' row must be a genuine measurement of the proposed architecture with text removed. The CTW1500 row (R=85.3, P=67.9, F=75.6) is identical, number-for-number, to TextSnake's published row in Table 3. Exact equality across all three metrics, including a precision of 67.9, is not a plausible outcome of an independent run of a different model; the simplest explanation is that this ablation row was not actually measured. The surrounding prose contains a concrete internal inconsistency: it describes the drop as '89.6% to 75.6%', but the full model F-score is listed as 90.2% elsewhere. This makes it unlikely that the numbers were freshly produced. If the no-text row is not real, the paper provides no valid evidence that textual features are essential, and the SOTA result cannot be attributed to the proposed semantic mechanism. I agree with the reader's weakest assumption; this is the same load-bearing weakness. The verdict REJECT is appropriate: the central architectural claim lacks a credible empirical basis. I see no code or checkpoints that would independently support the numbers, and the comparison tables also omit the closest cross-modal baselines, but the duplicated ablation alone is sufficient for rejection.","tokens_in":8701,"tokens_out":4867,"duration_ms":54222,"concrete_test":"Independently reproduce the 'No textual features' condition on CTW1500 with the paper's stated setting: frozen CLIP ResNet-50 image encoder, AFPN fusion, 3-layer vision-language decoder with the text branch and contrastive loss removed, 512x512 input, batch size 32, Adam lr=1e-4, 110 epochs, same augmentations. Evaluate at IoU 0.5 and compare the measured R/P/F to (85.3, 67.9, 75.6). If the re-run differs by more than rounding error—or if the authors cannot supply the no-text checkpoint or eval logs—Table 5 is not a genuine ablation and the central claim that textual features are essential is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that SAViL-Det's gains come from the language-vision decoder and text-to-pixel contrastive learning—rests on the ablation in Sec. 4.5 (Table 5). In the 'No textual features' row on CTW1500, the paper reports R=85.3, P=67.9, F=75.6. These are exactly the values reported for TextSnake [9] in Table 3, down to the unusual precision value. An independently measured run of the proposed CLIP+AFPN+3-layer-decoder model with text removed would not be expected to reproduce a different method's published scores exactly; this is strong evidence that the row was carried over from the baseline table rather than measured. The paragraph interpreting the ablation even cites a 14% drop from '89.6% to 75.6%', although the full model's F-score in Table 3 is 90.2%—another sign that the numbers are not internally consistent. If the no-text row is not a genuine measurement, then the only direct evidence that textual features are essential collapses, and the SOTA result is left without causal support. The exact match could in principle be a copy-paste error while the underlying run exists, but the burden is on the paper to show the run; no code, checkpoint, or logs are provided. On MLT2019, the no-text row (80.2/82.7/81.4) also lacks any independent verification, so the same concern applies to the paper's claim that language input is essential on both benchmarks.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAViL-Det, a scene text detection method that combines a frozen CLIP text encoder, a trainable ResNet-50 image encoder, an Asymptotic Feature Pyramid Network (AFPN), a transformer-based cross-modal decoder, and a text-to-pixel contrastive loss. The authors report state-of-the-art F-scores of 84.8% on MLT2019 and 90.2% on CTW1500, and they attribute these gains to the language-vision decoder and the text-to-pixel contrastive learning mechanism. The paper includes comparisons with prior methods, two ablation studies (decoder depth and removal of textual features), and qualitative examples.","tokens_in":8986,"tokens_out":5653,"duration_ms":66288,"significance":"The idea of injecting semantic text prompts into a multi-script text detector is timely and potentially useful, especially for handling curved and multilingual text where visual cues alone are insufficient. The paper is clearly organized and the experimental comparison covers a reasonable set of recent methods. However, as written, the method is not fully specified: the detection head and its training loss are never described, and the contrastive loss in Eq. (7) is not precisely defined. More seriously, the key ablation that is supposed to justify the textual component, Table 5, contains a row on CTW1500 that exactly duplicates the published TextSnake numbers from Table 3, and the accompanying text cites a full-model F-score of 89.6% that contradicts the 90.2% reported elsewhere. These issues mean that the central causal claim, that the language component is essential, is not currently supported by the evidence presented.","major_comments":[{"comment":"The detection head and the actual training objective for text detection are never described. The paper reports detection F-scores in Section 4.4, but no equation or text explains how the aligned feature map produced in Section 3.4 is converted into text instance polygons, nor what loss (segmentation, regression, or otherwise) supervises that prediction. Without this information, the proposed architecture is incomplete and the reported performance cannot be reproduced or attributed to the proposed components.","section":"Sections 3.3 and 3.4"},{"comment":"The sets P and N in the contrastive loss are not defined. The text says these are 'positive and negative pairs in the ground truth', but it never specifies which pixels are positive (all text pixels, per-instance pixels, or something else), how the single global text feature z_t is matched to multiple positive pixel features, or how negative pixels are sampled. Since the text-to-pixel contrastive learning is presented as a core contribution, this omission is load-bearing and prevents the reader from understanding or reimplementing the objective.","section":"Section 3.4, Eq. (7)"},{"comment":"In the CTW1500 ablation, the 'No textual features' row reports R=85.3, P=67.9, F=75.6, which is exactly the same R/P/F reported for TextSnake in Table 3, including the unusual precision value of 67.9. Moreover, the paragraph interpreting this ablation states the drop is from 89.6% to 75.6%, although the full model in Table 3 has F=90.2%. The exact match to a different method's published numbers strongly suggests that this row was carried over from the baseline table rather than being an independently measured variant of the proposed model. Since this ablation is the only direct evidence presented that textual features are essential, the central causal claim is unsupported.","section":"Section 4.5, Table 5 vs. Table 3"},{"comment":"The adaptive spatial fusion operation is written as F_v = α·F_v0 + β·F_v1 + γ·F_v3, with learnable weights α, β, γ, δ that sum to 1, but only three terms appear in the equation and δ is never defined, while F_v2 is omitted. This is inconsistent with Eqs. (1)–(3), where F_v1, F_v2, and F_v3 are computed sequentially. As written, the fusion mechanism is not well-defined and the reader cannot tell which features are actually combined.","section":"Section 3.2, Eq. (4)"},{"comment":"Model selection appears to be performed directly on the test sets. The decoder-depth ablation in Table 4 reports F-scores on the MLT2019 test set and the CTW1500 test set for depths 2 through 5, and the authors conclude that 3 layers is optimal based on these test-set numbers; no validation split or cross-validation is mentioned. Similarly, the three prompt templates in Table 1 are selected without any prompt-level ablation. Selecting hyperparameters on the test set inflates the reported results and weakens the validity of the state-of-the-art comparison, even though the final metrics themselves are computed against external benchmarks.","section":"Sections 4.3 and 4.5"}],"minor_comments":[{"comment":"The text says '1× upsampling' but F1 is at a lower resolution than F0, so the upsampling factor should be 2×; please correct this typo.","section":"Section 3.2, Eq. (1)"},{"comment":"The notation is inconsistent: the paper writes 'Let Ft ∈ R^(h×w)×d and Fv ∈ R^(N×C)', but earlier Ft is defined as R^(L×C) for text tokens and Fv as a visual feature map. Assigning spatial dimensions (h×w) to textual features is confusing; please unify the notation.","section":"Section 3.3"},{"comment":"The caption says 'the decoder takes visual features Ft' but Ft denotes textual features in the main text; the variable should be Fv or the caption should be corrected.","section":"Figure 3 caption"},{"comment":"There is a typo: 'descrease' should be 'decrease'.","section":"Section 4.5"},{"comment":"No code, checkpoints, or reproducibility statement is provided. Given the missing implementation details and the numerical inconsistency in Table 5, a reproducibility statement with links or logs is essential for any future revision.","section":"General"}],"recommendation":"reject","confidential_remarks":"The exact match between the 'No textual features' row in Table 5 and the TextSnake row in Table 3, together with the internal inconsistency about the 89.6% full-model F-score, goes beyond ordinary sloppiness. I am not accusing the authors of misconduct, but the evidence in the manuscript is consistent with the ablation row being copied from the baseline table rather than measured. Even setting that aside, the missing detection head and loss definition mean the method cannot be reproduced. I would not invite a revision unless the authors provide raw logs or code for the ablation and a complete specification of the training objective; as it stands, the paper's central claim is not verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zighem and Hadid's SAViL-Det hits the two things that matter in scene text detection — a plausible idea and decent headline numbers — but the experimental support for the central claim doesn't hold up. The architecture genuinely recombines CLIP, AFPN, a transformer decoder with cross-attention, and a pixel-level contrastive loss, and the decoder-depth ablation is reported cleanly across five IoU thresholds. That part is worth something.\n\nThe problem is the 'No textual features' ablation in Table 5. On CTW1500 it reports R=85.3, P=67.9, F=75.6 — exactly TextSnake's published numbers from Table 3 in the same paper, down to the decimal. An independent run of their own model with text removed landing on another method's exact scores is not credible. The paragraph interpreting it then says the drop is from 89.6% to 75.6%, but their full model's F-score in Table 3 is 90.2% (89.6 is SRFormer). So the ablation text is internally inconsistent too. Without that row, the only direct evidence that language input is essential collapses, and the SOTA claim loses its causal story.\n\nSmaller issues accumulate: the detection head and its loss are never described, Eq. (7)'s positive/negative pixel sets are undefined, the closest cross-modal baselines (e.g., the CLIP-based detector they cite as [21], and the weakly-supervised VLP method [17]) don't appear in the comparison tables, and the prompt/decoder-depth choices look like they were tuned on the test sets. No code or checkpoints are provided, which makes the duplicated row harder to dismiss.\n\nThe idea is not bad; the execution and reporting are. I would not send this to peer review in its current form. The authors should be asked to release code/checkpoints and rerun the no-text ablation, plus fill in the detection head and loss details. If the rerun reproduces the reported full-model numbers and shows the language input actually matters, the paper could come back as a competently executed incremental method. As is, it's a desk reject.","headline":"The architecture idea is plausible, but the 'no text' ablation on CTW1500 exactly duplicates TextSnake's published numbers, so the paper's central claim lacks credible support as it stands.","tokens_in":9537,"tokens_out":3153,"would_cite":false,"duration_ms":34930,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A text detector that feeds CLIP language prompts into pixel-level features reports top F-scores of 84.8% on MLT-2019 and 90.2% on CTW1500.","keywords":["scene text detection","vision-language model","CLIP","multi-lingual text","curved text","cross-modal attention","contrastive learning"],"falsifier":"Re-run the 'No Textual Features' configuration on CTW1500 with the same training data and schedule; if the measured F-score is not close to the reported 75.6, the 14.6-point gap attributed to language input is not supported.","tokens_in":8463,"feed_emoji":"🔤","tokens_out":9518,"duration_ms":105313,"temperature":0.7,"pith_summary":"SAViL-Det is a proposal for detecting text in natural images by treating the text prompt as a first-class input rather than an optional helper. The paper claims that a decoder which first applies self-attention to visual features and then cross-attention with features from a frozen CLIP text encoder lets semantic information from the prompt propagate into pixel-level features, and that a text-to-pixel contrastive loss makes those features align with the prompt. On two benchmark settings the paper reports F-scores of 84.8% on the multi-lingual MLT-2019 set and 90.2% on the curved-text CTW1500 set, ahead of the comparison methods in its tables. This matters because multi-script and curved text are exactly the cases where visual shape alone is ambiguous, so a detector that can use language context promises to be more reliable in the wild.","feed_headline":"Prompt-fed text detector tops curved and multi-script tests","feed_subtitle":"Injecting text semantics into pixels lifts F-scores to 84.8% and 90.2% on two hard benchmarks.","key_machinery":"The load-bearing component is the language-vision decoder: a transformer layer stack in which multi-head self-attention runs over visual features, multi-head cross-attention takes keys and values from the frozen CLIP text tokens, and a feed-forward MLP with residual connections outputs multi-modal features. Around it sit the AFPN multi-scale fusion, which combines non-adjacent feature levels with learnable spatial weights, and a text/image projector that upsamples the decoder output to image resolution and applies a sigmoid contrastive loss between the global text vector and each pixel vector. Together these make the prompt's semantic content able to reshape pixel-level features before the detection head.","core_discovery":"The paper's central claim is that textual semantics are not just side information but the active ingredient that improves detection. Concretely, SAViL-Det encodes three script-agnostic prompts with the frozen CLIP text encoder, fuses multi-scale visual features from a CLIP ResNet-50 backbone with an Asymptotic Feature Pyramid Network, and then runs a decoder whose cross-modal attention lets each visual token attend to text-token features. A contrastive loss over the global text embedding and projected pixel embeddings pushes text pixels close to the prompt and background far from it. The paper reports that this pipeline reaches 84.8% F on MLT-2019 and 90.2% F on CTW1500, and that removing the textual features lowers the F-score by 3.4 points on MLT-2019 and from 90.2 to 75.6 on CTW1500.","pith_inferences":["A natural next test, not run in the paper, is to change the prompt's script names or replace them with nonsense words; if the F-score moves little, the gain might come from conditioning in general rather than from semantic script knowledge.","The contrastive loss treats text-to-pixel similarity with a single sigmoid; the paper does not specify how negative pairs are sampled, so varying the negative sampling rule (other text instances vs background) is a testable knob that could explain the CTW1500 precision jump.","If the decoder truly propagates token-level semantics, the architecture should extend to text spotting (reading the detected words) by conditioning on the actual transcription instead of generic prompts, a possibility the paper lists as future work.","The surprising CTW1500 no-text result, which matches TextSnake's published row digit for digit, should make readers treat the ablation gap as provisional rather than final; reproducing that row is the cheapest way to stress-test the main claim."],"forward_implications":["If the reported scores hold, a single prompt-conditioned model can outperform dedicated multi-lingual and curved-text detectors without per-script specialisation.","The decoder-depth ablation indicates the benefit comes from exactly three transformer layers; going to four or five layers degrades F-scores on both datasets.","The prompt-ablation numbers imply that language input is worth 3.4 F-points on MLT-2019 and about 14.6 points on CTW1500, meaning curved text is where semantic guidance pays off most.","Because the prompt list names the scripts to detect, the same model outputs script labels as well as boxes on MLT-2019."],"supporting_citations":[{"why":"Supplies the frozen CLIP image and text encoders that all visual and textual features come from.","marker":"[12]"},{"why":"Defines the AFPN fusion that combines multi-scale visual features before the decoder.","marker":"[18]"},{"why":"Provides MLT-2019, the multi-lingual benchmark where the paper reports 84.8% F-score.","marker":"[10]"},{"why":"Provides CTW1500, the curved-text benchmark where the paper reports 90.2% F-score.","marker":"[8]"},{"why":"MixNet is the closest comparison on CTW1500 that SAViL-Det reports slightly surpassing.","marker":"[22]"},{"why":"TextSnake's published F=75.6 row is the exact set of numbers reported for the no-text ablation, so it is the key reference for checking that ablation.","marker":"[9]"}],"fun_headline_variants":["Text prompts sharpen detection on curved and multilingual scenes","Semantics injected into pixels improve text detection","Language guidance boosts multi-script and curved text detection","CLIP-based model reads text by aligning words to pixels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 'No Textual Features' row was actually measured rather than carried over from a published baseline, since on CTW1500 its recall, precision, and F-score are exactly TextSnake's published numbers.","fun_headline_variants_meta":{"raw":{"variants":["Text prompts sharpen detection on curved and multilingual scenes","Semantics injected into pixels improve text detection","Language guidance boosts multi-script and curved text detection","CLIP-based model reads text by aligning words to pixels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00052,"raw_usage":{"total_tokens":2505,"prompt_tokens":920,"completion_tokens":1585,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":1524}},"tokens_in":536,"tokens_out":1585,"duration_ms":12931,"temperature":1.0,"reasoning_tokens":1524,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:41:26.048509+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the 'No Textual Features' configuration on CTW1500 with the same training data and schedule; if the measured F-score is not close to the reported 75.6, the 14.6-point gap attributed to language input is not supported.","supporting_citations":[{"cited_title":"In: ICML 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen CLIP image and text encoders that all visual and textual features come from."},{"cited_title":"In: 2023 IEEE International Conference on Systems, Man, and Cybernetics (SMC)","cited_arxiv_id":null,"evidence_quote":"Defines the AFPN fusion that combines multi-scale visual features before the decoder."},{"cited_title":"In: ICDAR","cited_arxiv_id":null,"evidence_quote":"Provides MLT-2019, the multi-lingual benchmark where the paper reports 84.8% F-score."},{"cited_title":"Pattern Recognition90 (2019)","cited_arxiv_id":null,"evidence_quote":"Provides CTW1500, the curved-text benchmark where the paper reports 90.2% F-score."},{"cited_title":"In: ECCV","cited_arxiv_id":null,"evidence_quote":"TextSnake's published F=75.6 row is the exact set of numbers reported for the no-text ablation, so it is the key reference for checking that ablation."}],"review_version":1}