{"id":"762a23d5-1b56-4426-b241-87a47834d59c","arxiv_id":"2506.06966","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"NationalCSL-DP provides dual-view RGB videos for the full Chinese national sign vocabulary and shows simple score-level fusion improves isolated sign recognition.","lead":"This paper introduces NationalCSL-DP, a new Chinese sign language dataset with 134,140 videos of 6,707 words filmed from front and left views. It also presents a CNN-transformer baseline and shows that fusing the two views at prediction time improves isolated sign recognition accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never specifies the train/val/test split protocol; with only 10 signers, a random video-level split would leak signer identity and inflate the reported dual-view and fusion gains.","rationale":"The paper's primary contribution is a new benchmark dataset, and its usefulness depends on a valid evaluation protocol. All reported accuracy numbers and comparative conclusions—single-view vs dual-view, early/late/plus fusion—derive from a split that is never described in the manuscript. This is an internal reproducibility and validity issue, not a disagreement with external consensus: with only 10 signers, signer identity is a trivially available shortcut, and a video-level random split would invalidate the claim that accuracy reflects sign-language generalization. The downloadable partition files make this concern directly checkable, but the paper should state the protocol explicitly. I do not find a more load-bearing internal inconsistency: the method equations appear standard aside from a likely typo in Eq. (2), where p_left should be computed from X_left rather than X_front, and a missing '+' symbol in Eq. (8). The plus-fusion gain may in part reflect simple ensembling of two independently trained models, but the evaluation-protocol gap is the more fundamental issue because it conditions every empirical claim in the paper. Since the reader already returned a CONDITIONAL verdict based on substantially the same concern, my stress-test does not change the recommended verdict.","tokens_in":13488,"tokens_out":5313,"duration_ms":61909,"concrete_test":"Download the published NationalCSL-DP partition files and compute the overlap of signer identities between the training and test partitions. If any signer appears in both, rerun the plus-fusion experiment with a signer-disjoint split (e.g., train on 8 signers, test on 2); if the top-1 accuracy on NationalCSL6707 drops substantially or the 5.27% plus-fusion gain over the front view shrinks, the reported numbers are partly due to identity leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The empirical claims—including that dual-view input is beneficial and that plus fusion 'significantly increases' accuracy—all depend on the evaluation split, but the manuscript never states it. Section 3.2 only says that K glosses were 'randomly selected' to form the five subsets and that 'partition files' are downloadable; it does not say whether the train/val/test split is signer-independent, gloss-disjoint, or even what fraction of videos goes into each partition. Section 5.2 mentions early stopping on validation loss but does not define the validation set. With exactly 10 signers in NationalCSL-DP, a random video-level split would place the same signer in both training and test sets. A ResNet-34 spatial encoder can then exploit per-signer appearance (face, body, clothing, studio background) instead of sign movement, inflating top-1 accuracy. Because the same ambiguous protocol underlies all comparisons in Tables 1 and 2, the central claim that multisview fusion improves sign recognition is not established as signer-independent generalization. The downloadable partition files may resolve this, but as written the text leaves the protocol unstated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces NationalCSL-DP, a dual-view Chinese isolated sign language recognition (ISLR) dataset with 134,140 RGB videos from 10 signers covering 6,707 glosses from the Chinese National Sign Language vocabulary, captured simultaneously from front and left views. The dataset is organized into five subsets (200, 500, 1000, 2000, and 6707 glosses). The authors also propose a CNN-transformer baseline and compare early, late, and plus fusion strategies, reporting that dual-view input and plus fusion improve top-1 accuracy, reaching 69.61% on the 6707-gloss subset. The paper claims that this is the first ISLR dataset covering the full national vocabulary and providing aligned dual-view videos.","tokens_in":13716,"tokens_out":3984,"duration_ms":45129,"significance":"If the dataset and the reported results hold up, NationalCSL-DP is a potentially valuable community resource: it is larger in vocabulary than existing Chinese ISLR datasets, it provides two synchronized views to address hand occlusion, and it is released with partition files under a CC-BY license. The proposed plus-fusion baseline is simple and could serve as a useful reference. The paper also has several strengths that should be credited: the dataset construction was supervised by sign-language experts, cross-verification was performed, and the authors provide a concrete download location. However, the empirical claims in the paper are currently not fully supported because the evaluation protocol is underspecified and no uncertainty quantification is given. The missing train/test split description is the most serious issue, since all conclusions about dual-view fusion depend on it.","major_comments":[{"comment":"The train/validation/test split protocol is never specified. Section 3.2 states only that K glosses were randomly selected to form the five subsets and that partition files are downloadable; it does not state whether the video-level splits are signer-independent, whether paired front/left videos of the same gloss are kept in the same partition, or what proportion of videos is assigned to training, validation, and test. Section 5.2 mentions early stopping on validation loss but does not define the validation set. With only 10 signers, a random video-level split could put the same signer in both training and test sets, allowing the model to exploit per-signer appearance rather than sign content, which would inflate the dual-view and fusion gains reported in Tables 1 and 2. The authors should state the exact split protocol and, if the splits are not signer-independent, rerun the evaluation under a signer-independent protocol.","section":"Section 3.2 and Section 5.2"},{"comment":"All results are reported as single numbers with no error bars, no repeated trials, and no statistical significance test. The text repeatedly says that dual-view input and plus fusion 'significantly increase' accuracy, but on NationalCSL6707 the dual-view over front-view top-1 improvement is 3.26% for SL-GCN and 5.27% for CNN-transformer (Table 1). These differences could be within run-to-run variability for a single seed and a fixed hyperparameter configuration. The authors should report mean and standard deviation over multiple runs, or at least provide a clear statement about stochasticity, and soften the word 'significantly' unless a statistical test is supplied.","section":"Tables 1 and 2, Section 5.3 and 5.4"},{"comment":"The claim that the CNN-transformer is a 'strong baseline' is not externally anchored. The only comparison is with SL-GCN on the authors' own new dataset, so there is no previously published result on NationalCSL-DP to calibrate what 'strong' means. This is not by itself a flaw in a dataset paper, but the wording overclaims. The authors should either compare with additional established ISLR methods adapted to this dataset or explicitly limit the claim to 'a baseline for NationalCSL-DP' without the 'strong' qualifier.","section":"Section 5.3"}],"minor_comments":[{"comment":"The number of words in the Chinese National Sign Vocabulary is given as 8124 in the Abstract and Introduction but as 8214 in Section 3.1; please reconcile these numbers.","section":"Section 3.1"},{"comment":"Equation (2) defines p_left = S_left(X_front), but it should presumably be S_left(X_left); this typo makes the notation confusing.","section":"Equation (2)"},{"comment":"There are two subsections numbered 5.1 ('Experimental design' and 'Evaluation metric'); the second should be renumbered.","section":"Section 5.1"},{"comment":"Equation (8) is missing the fusion operator between M_front(X_front) and M_left(X_left), and it is unclear whether these outputs are logits or probabilities before the Softmax; please make this explicit.","section":"Equation (8)"},{"comment":"The video count is inconsistently reported as 134,140 in the Abstract but 131K+ in the Introduction, Contributions, and Conclusions; please standardize.","section":"Abstract and Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The dataset itself appears to be a worthwhile contribution, and the paper is generally readable. The main barrier is the missing evaluation protocol: without a clear signer-independent split description, the headline numbers cannot be trusted. The downloadable partition files may resolve this, but as written the manuscript does not provide the necessary information. I would encourage the editor to ask for the split specification and, if needed, a re-evaluation under a signer-independent protocol before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The dataset is the real contribution here. NationalCSL-DP is the first dual-view RGB dataset covering the full Chinese national sign vocabulary after deduplication (6,707 glosses, 134k+ videos, front and left views from 10 signers), and that alone makes the paper worth engaging with. The recording was supervised by sign language experts, the annotation was cross-checked, and the release under CC-BY with partition files is the right way to do this. Credit where it is due: this fills a genuine gap in Chinese ISLR resources.\n\nThe baseline itself is standard: ResNet-34 per frame plus a transformer over time, with early, late, and plus fusion. The plus fusion is just an ensemble of the two view-specific softmax outputs, which explains why it helps while early and late fusion do not. That is a fine baseline result. The authors are honest that they do not analyze why early/late fusion fail.\n\nThe soft spots are real, though. Most importantly, the paper never specifies the train/val/test split protocol. Section 3.2 only says glosses were randomly selected for the five subsets; it does not say whether videos are split by signer, by gloss, or randomly at the video level. With only 10 signers, a random video-level split would place the same signer in both train and test, letting the model exploit per-signer appearance rather than sign movement. That would inflate the reported dual-view gains. The downloadable partition files may resolve this, but as written the claim that dual-view fusion improves generalization is not established. This is the load-bearing issue. Secondary issues: no error bars or repeated trials, no comparison against external benchmarks, and no code release for the baseline. Minor inconsistencies: the abstract and conclusion both mention 8,124 sign words, while the deduplicated vocabulary is 6,707; equation (2) has a typo where the left-view features are computed from the front-view input.\n\nOverall, the dataset deserves serious consideration, and the paper is suitable for peer review, but the revision must describe the split protocol precisely, provide signer-independent results if that is what was done, and ideally add variance estimates or external validation. If the split is signer-independent, the main conclusion likely holds; if not, the numbers need re-evaluation. I would take this to a reading group for the dataset design, and I would cite it if I worked on sign language recognition.","headline":"The NationalCSL-DP dataset is a genuinely valuable resource, but the paper's empirical claims are under-specified because the train/test split protocol is never stated.","tokens_in":14235,"tokens_out":1490,"would_cite":true,"duration_ms":17983,"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's central claim is that a dual-view dataset covering the full Chinese national sign vocabulary, combined with a simple prediction-time fusion of two independently trained view models, substantially improves isolated sign…","keywords":["isolated sign language recognition","dual-view dataset","Chinese sign language","CNN-transformer","feature fusion","NationalCSL-DP","hand occlusion"],"falsifier":"Run the plus-fusion CNN-transformer on a signer-disjoint split, training on eight signers and testing on the two unseen signers, and compare dual-view against single-view top-1 accuracy; if the dual-view advantage shrinks or disappears, the claim that the dual-view dataset and plus fusion drive the accuracy gain is not established.","tokens_in":13302,"feed_emoji":"🤟","tokens_out":5567,"duration_ms":53409,"temperature":0.7,"pith_summary":"The paper presents NationalCSL-DP, a dual-view Chinese isolated sign language recognition dataset containing 134,140 videos of 6,707 glosses recorded from front and left cameras, and claims it is the first dataset to cover the complete Chinese national sign vocabulary while providing aligned two-view videos. It further argues that feeding both views to a recognition model improves accuracy over either single view, and that the best way to exploit the two views is a 'plus fusion' strategy: train a front-view model and a left-view model separately, then sum their predictions at inference. On the largest subset the proposed CNN-transformer baseline reaches 69.61% top-1 accuracy, whereas early and late fusion of the two views fail to beat single-view performance. A sympathetic reader would care because the dataset supplies a vocabulary-complete, occlusion-relevant benchmark for sign language recognition and a strong baseline to compare against.","feed_headline":"Dual-view videos lift Chinese sign recognition to 69.6%","feed_subtitle":"New dual-view dataset spans all 6,707 national glosses; a simple view-fusion baseline beats single-view models.","key_machinery":"The load-bearing mechanism is the dataset-plus-fusion pipeline. NationalCSL-DP was built by deduplicating the 8,124-word Chinese national sign vocabulary to 6,707 distinct glosses (dropping glosses with identical hand motions, and tagging homographs and regional variants with suffixes such as '1-1' and '2-1'), then recording ten fluent signers with two synchronized RGB cameras at 1920×1080 and 50 fps from the front and left sides. The recognition baseline extracts per-frame 512-dimensional features with a ResNet-34, models temporal dependencies with a 4-layer, 8-head transformer encoder, and predicts through an MLP head. 'Plus fusion' trains two such models independently, one per view, and combines their softmax outputs by addition before the final argmax; this is the operation that yields the paper's best results.","core_discovery":"The central discovery is that dual-view sign videos are genuinely more informative than single-view videos, but only when the model is allowed to learn each view's temporal dynamics independently and fuse at the decision level. The paper shows this by comparing a skeleton-based model (SL-GCN) and its own CNN-transformer on front-only, left-only, and dual-view inputs across five subsets (200, 500, 1000, 2000, and 6707 glosses). Dual-view plus fusion raises top-1 accuracy on the 6707-gloss subset to 69.61%, up from 64.34% for the front view and 40.99% for the left view with the same backbone; consistent gains appear on every subset. The paper also reports that early fusion (concatenating per-frame CNN features) and late fusion (concatenating transformer outputs) do not improve over single-view accuracy, and on the largest subset they actually hurt, which the authors leave for future work.","pith_inferences":["If the dataset's split is signer-disjoint and the gains persist, NationalCSL-DP becomes a natural testbed for occlusion-robust sign recognition, since the left view often reveals hands hidden in the front view.","The failure of early and late fusion hints that the two views' temporal alignments differ; frame-wise feature concatenation may be mismatched. Cross-view attention or view-alignment pretraining is a testable extension the paper does not explore.","Because the vocabulary list is the deduplicated national standard and preserves regional variants via suffixes, the dataset could support building sign-language dictionaries and dialect-aware recognition systems, not just isolated word benchmarks.","A natural next step is continuous sign language recognition: if sentence-level models also benefit from dual-view plus fusion, the dataset's design could extend beyond isolated glosses."],"forward_implications":["Dual-view input improves top-1 ISLR accuracy for both RGB and skeleton-based models across all five dataset sizes.","Plus fusion consistently outperforms early fusion and late fusion, making it a simple and effective baseline for two-view sign recognition.","Early and late fusion of dual-view features can underperform single-view input, especially on large vocabularies, indicating that naive feature concatenation does not capture complementary view information.","The five-scale dataset arrangement (200 to 6707 glosses) supports studying how recognition difficulty scales with vocabulary size.","The CNN-transformer with plus fusion is offered as a strong baseline for future work on NationalCSL-DP."],"supporting_citations":[{"why":"DEVISIGN, the largest prior Chinese ISLR dataset with 2,000 words; supplies the coverage gap the new dataset fills.","marker":"[7]"},{"why":"SL-GCN (Jiang et al.), the skeleton-based comparison baseline used to show dual-view benefits generalize across model types.","marker":"[19]"},{"why":"WLASL, whose subset arrangement (200, 500, 1000, 2000 glosses) the paper follows for NationalCSL-DP's five scales.","marker":"[25]"},{"why":"AUTSL, a multi-modal Turkish dataset whose depth data address occlusion, used as comparison for single-view limitations.","marker":"[26]"},{"why":"NMFs-CSL, the existing 2,000-gloss Chinese dataset recorded by 10 signers, providing the closest prior CSL benchmark.","marker":"[27]"},{"why":"MM-WLAuslan, a recent multi-view multi-modal dataset that also targets occlusion but does not cover the full vocabulary.","marker":"[29]"}],"fun_headline_variants":["Dual-view input raises sign accuracy to nearly 70%","CNN-transformer with dual views hits 69.6% on sign language","View fusion pushes Chinese ISLR past 69%","Two camera angles improve sign recognition by 5 points","Dual-view dataset lifts sign accuracy to 69.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes its train/test split is fair (for example, signer-independent), but it never describes how the videos were divided, so if the same signer appears in both training and testing the reported dual-view and fusion gains could partly come from signer memorization.","fun_headline_variants_meta":{"raw":{"variants":["Dual-view input raises sign accuracy to nearly 70%","CNN-transformer with dual views hits 69.6% on sign language","View fusion pushes Chinese ISLR past 69%","Two camera angles improve sign recognition by 5 points","Dual-view dataset lifts sign accuracy to 69.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2731,"prompt_tokens":992,"completion_tokens":1739,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":1653}},"tokens_in":608,"tokens_out":1739,"duration_ms":14136,"temperature":1.0,"reasoning_tokens":1653,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:44:30.377559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the plus-fusion CNN-transformer on a signer-disjoint split, training on eight signers and testing on the two unseen signers, and compare dual-view against single-view top-1 accuracy; if the dual-view advantage shrinks or disappears, the claim that the dual-view dataset and plus fusion drive the accuracy gain is not established.","supporting_citations":[{"cited_title":"ACM Transactions on Accessible Computing, 8(4): 1-21","cited_arxiv_id":null,"evidence_quote":"DEVISIGN, the largest prior Chinese ISLR dataset with 2,000 words; supplies the coverage gap the new dataset fills."},{"cited_title":"In Procee dings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'21), pp 3413-3423","cited_arxiv_id":null,"evidence_quote":"SL-GCN (Jiang et al.), the skeleton-based comparison baseline used to show dual-view benefits generalize across model types."},{"cited_title":"In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV'20), pp 1459-1469","cited_arxiv_id":null,"evidence_quote":"WLASL, whose subset arrangement (200, 500, 1000, 2000 glosses) the paper follows for NationalCSL-DP's five scales."},{"cited_title":"IEEE Access, 8: 181340-181355","cited_arxiv_id":null,"evidence_quote":"AUTSL, a multi-modal Turkish dataset whose depth data address occlusion, used as comparison for single-view limitations."},{"cited_title":"ACM transactions on multimedia com puting, communications, and applications, 17(3): 1-19","cited_arxiv_id":null,"evidence_quote":"NMFs-CSL, the existing 2,000-gloss Chinese dataset recorded by 10 signers, providing the closest prior CSL benchmark."},{"cited_title":"In Proceedings of the Ann ual Conference on Neural Information Processing Systems (NeurIPS’24)","cited_arxiv_id":null,"evidence_quote":"MM-WLAuslan, a recent multi-view multi-modal dataset that also targets occlusion but does not cover the full vocabulary."}],"review_version":1}