{"id":"edd692e9-38d2-4330-8485-f2e66cfa40e4","arxiv_id":"2502.02196","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An ensemble of three Video Swin Transformer sizes with RGB and depth fusion achieves 20.29% RGB and 24.53% RGB-D top-1 accuracy, ranking third in the CV-ISLR challenge.","lead":"This paper describes a competition entry that averages three sizes of Video Swin Transformer models, plus a fusion of RGB and depth branches, for recognizing sign language from side camera angles. It ranked third in both CV-ISLR tracks, but its accuracy (20.29% RGB, 24.53% RGB-D) is far below the winning entries (56.87% and 57.97%).","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RGB-D ensemble gain lacks a single-model baseline: Table 2 only establishes the ensemble benefit for the RGB track, not for RGB-D fusion.","rationale":"The paper is a competition report: it compares three model sizes, shows that a weighted average helps on RGB, and reports a final RGB-D number. The reader's conditional verdict is reasonable. The most load-bearing gap is not weight sensitivity per se, but the absence of any single RGB-D baseline, which makes the cross-modal ensemble component untestable from the reported tables. The RGB-track ablation is legitimate evidence and should be credited. The large gap to the top-ranked teams (56.87% vs. 20.29%) weakens the 'competitive performance' wording, but it does not invalidate the narrow internal comparison. The code link is unverified; re-running the RGB ablation and adding single RGB-D baselines would settle the main question, and the proposed test is straightforward relative to the training already performed.","tokens_in":8225,"tokens_out":6703,"duration_ms":68621,"concrete_test":"Train and evaluate the three RGB-D VST variants (Small/Base/Large) on the same test set, using the fusion defined by Eq. (7), and report their individual Acc@1 values alongside the 24.53% ensemble. Then sweep the ensemble weights over a coarse grid (e.g., all triples summing to 1 in 0.1 increments) on a held-out validation split. If a single RGB-D VST already approaches 24.53%, the multimodal ensemble contributes little beyond modality fusion; if the accuracy at the reported weights is a narrow peak, the hand-set weights are overfit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ensemble learning drives the reported gains is only directly supported for the RGB track. Table 2 compares the RGB ensemble (20.29%) against VST-Small/Base/Large (14.84%/17.51%/17.04%), so the RGB improvement is internally consistent. For the RGB-D track, however, the table reports only the final ensemble accuracy (24.53%) and no accuracy for any single RGB-D model. The 4.24-point gap between the RGB ensemble (20.29%) and the RGB-D ensemble (24.53%) can therefore be explained entirely by the addition of the depth modality, rather than by multi-dimensional ensembling. Section 3.3.2, Eq. (7), defines Lambda_rgbd as a sum over model sizes of Phi(Z_r oplus Z_d), implying that individual RGB-D variants exist and should be reported as baselines. Without those entries, the statement in Section 4.2 that 'the ensemble learning strategy effectively captures the strengths of models with different dimensions' is not verified for the multimodal track. The hand-set weights (lambda_l=0.4, lambda_b=0.4, lambda_s=0.2; lambda_r=0.65, lambda_d=0.35) also lack sensitivity analysis, but that is secondary: even with robust weights, the RGB-D comparison needs a single-model baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper describes the authors' third-place solution to the WWW 2025 Cross-View Isolated Sign Language Recognition (CV-ISLR) challenge on the MM-WLAuslan dataset. The method trains Video Swin Transformer models in three sizes (Small, Base, Large) on RGB and depth videos, combines the outputs within each modality by a weighted sum of predictions, and then fuses the RGB and depth branches by another weighted sum. The paper reports top-1 accuracies of 20.29% for the RGB track and 24.53% for the RGB-D track, compared with a best single RGB model accuracy of 17.51%. The central claim is that the ensemble strategy captures complementary strengths of models with different capacities and improves cross-view robustness.","tokens_in":8623,"tokens_out":5334,"duration_ms":51888,"significance":"If the reported gains are robust, the paper provides a useful empirical data point on ensembling Video Swin Transformers for cross-view sign language recognition, and the public code repository is a practical strength for reproducibility. The RGB-track result is internally consistent: the ensemble of VST-Small/Base/Large improves over every single RGB model. However, the novelty is limited to an engineering combination of existing components, and the experimental evidence is incomplete for the RGB-D track and for the stability of the hand-set ensemble weights, so the significance as currently presented is modest.","major_comments":[{"comment":"The central claim in Section 4.2 that 'the ensemble learning strategy effectively captures the strengths of models with different dimensions' is directly supported only for the RGB track. Table 2 reports VST-Small/Base/Large accuracies for RGB (14.84%/17.51%/17.04%) and for depth (14.01%/16.46%/17.58%), but for the RGB-D track it reports only the ensemble accuracy (24.53%) and leaves the corresponding single-model entries empty. Since Eq. (7) defines the RGB-D ensemble as a combination over model sizes of fused RGB-D models, individual fused models must exist; without reporting at least one single RGB-D baseline, the 4.24-point gap between the RGB ensemble (20.29%) and the RGB-D ensemble (24.53%) could be explained entirely by the additional depth modality rather than by ensembling. Please add per-size fused-model results, or at minimum a VST-Large RGB-D baseline.","section":"Section 4.2, Table 2"},{"comment":"The ensemble weights are hand-set (lambda_l:lambda_b:lambda_s = 0.4:0.4:0.2 and lambda_r:lambda_d = 0.65:0.35) with no sensitivity analysis or selection criterion. The reported RGB improvement (20.29% vs. 17.51%) compares a weighted combination against individual models; if these weights were tuned on the validation set, the gain could reflect weight selection rather than a robust property of ensembling. Please report a small sweep over weights (e.g., uniform averaging, majority vote, and neighboring weight vectors) or otherwise justify that the result is insensitive to the choice.","section":"Section 4.1, Implementation Details"},{"comment":"No repeated runs, seeds, or error bars are reported in Table 2. The differences among single models are small (VST-Base 17.51% vs. VST-Large 17.04%), and the ensemble gain over VST-Base is 2.78 percentage points; without an estimate of run-to-run variance it is difficult to assess whether this gain is statistically meaningful. If the challenge test set allows only a single evaluation, please state this explicitly and report validation-set results with mean and standard deviation over seeds for the main configurations.","section":"Section 4.1 / Table 2"}],"minor_comments":[{"comment":"The summation index i does not appear in the summands of Eqs. (5) and (6); as written, each term is identical. The intended meaning is presumably Phi_i(Z_r) and Phi_i(Z_d) for i in {s,b,l}. Please correct the notation.","section":"Section 3.3.1, Eqs. (5)-(6)"},{"comment":"Eq. (7) also contains an undefined symbol omega^(i)_rgbd and a summand Phi(Z_r oplus Z_d) that does not depend on i. Please clarify how the fused RGB-D models are indexed and how their parameters enter the sum.","section":"Section 3.3.2, Eq. (7)"},{"comment":"The sentence 'For the VST model, Large, Base and Small sizes of C correspond to 96, 128, and 192' appears to reverse the usual mapping (Large has the largest C). Please verify and state the actual channel widths used for each model size, since the multi-dimensionality claim depends on these settings.","section":"Section 4.1, Implementation Details"},{"comment":"The depth input is described as Z_d in R^{T x H x W x 3}; since depth maps are single-channel, please clarify how the three channels are formed (e.g., replication or colorization).","section":"Section 3.1, Task Definition"},{"comment":"The missing entries in Table 2 (depth ensemble and RGB-D per-backbone rows) should at least be marked with a dash and explained in the caption, so the reader can see which configurations were not evaluated.","section":"Table 2"},{"comment":"The term 'competitive performance' is not qualified: with a gap of more than 36 points to the first-place RGB result, the claim should be explicitly limited to ranking among submitted challenge systems rather than implying accuracy parity with state-of-the-art methods.","section":"Abstract / Section 4.2"},{"comment":"References [22] and [23] are the same paper (Auslan-Daily) listed twice with different years; please deduplicate and use the correct venue/year.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is a challenge-report paper whose main empirical claim is plausible but incompletely supported: the missing RGB-D single-model baseline is a load-bearing gap, and the hand-set ensemble weights need a sensitivity check. Both are straightforward to address within the manuscript's scope, so I recommend major revision rather than rejection. I would also ask the authors to correct the channel-width sentence and the equations in Section 3.3 before publication, as they currently impede formal reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a WWW 2025 challenge report that took third place in both CV-ISLR tracks by averaging predictions from three Video Swin Transformer sizes and then fusing RGB and depth. The only genuinely established result is that the RGB ensemble beats each single RGB model (20.29% vs 17.51% best single). Everything else is presented too quickly or overclaimed.\n\nWhat's good: the paper is short, honest about the gap to the winners (20% vs 57% and 40%), and the RGB ablation in Table 2 directly supports the narrow ensemble claim. The code link is provided. That is a useful data point for anyone working on cross-view sign language recognition.\n\nThe soft spots are real. The most important one is that Table 2 has no single-model RGB-D baseline. The 24.53% RGB-D number could be almost entirely due to adding the depth stream, not to ensembling the three model sizes. Eq. 7 implies individual RGB-D models exist, so leaving them out is a genuine omission, not a cosmetic one. The hand-set weights (0.4/0.4/0.2 and 0.65/0.35) also lack any sensitivity analysis, though that is secondary to the missing baseline. There are also no error bars, and the equations in Section 3.3 don't actually show the weights that are described in the implementation details. The phrase \"competitive performance\" in the abstract is doing too much work when the accuracy is a third of the winner's.\n\nI don't think the central RGB claim is in doubt. But the RGB-D story is undersupported as written, and the conclusion overgeneralizes. The paper is coherent, not incoherent; it is a thin but honest report with one clear gap.\n\nWho this is for: someone organizing or entering a sign-language challenge will find the ablation useful as a simple baseline. A general CV audience won't learn anything new. As a peer-reviewed paper, it needs the missing baseline and a sensitivity analysis before the RGB-D claim is supportable.\n\nMy recommendation: send it to peer review only if the venue tolerates workshop-style challenge reports and the authors add the missing RGB-D baselines. As is, a serious editor could either desk reject for low novelty or send it out for a quick revision. I'd lean toward sending it out because the missing experiment is cheap and the question is empirically testable.","headline":"A third-place challenge report that shows a real RGB ensemble gain but leaves the RGB-D claim undersupported; worth a quick read for the ablation, but not a strong paper.","tokens_in":9042,"tokens_out":3116,"would_cite":false,"duration_ms":29335,"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":"A two-stage ensemble of Video Swin Transformer sizes (Small, Base, Large) raises cross-view sign recognition to 20.29% top-1 on RGB and 24.53% on RGB-D on MM-WLAuslan.","keywords":["cross-view sign language recognition","ensemble learning","Video Swin Transformer","multi-modal fusion","RGB-D recognition","MM-WLAuslan","WWW 2025 challenge","top-1 accuracy"],"falsifier":"Re-score or retrain the ensemble over a grid of weights—for example, $\\lambda_l$ and $\\lambda_b$ from 0 to 1 in steps of 0.1 with $\\lambda_s = 1-\\lambda_l-\\lambda_b$, and $\\lambda_r$ from 0 to 1 in steps of 0.1—and check whether the reported 20.29% RGB and 24.53% RGB-D top-1 accuracies lie on a broad plateau or a narrow peak; a narrow peak would indicate the ensemble gain is a weight-selection artifact, not a general property of capacity diversity.","tokens_in":8015,"feed_emoji":"🤟","tokens_out":10467,"duration_ms":87087,"temperature":0.7,"pith_summary":"The paper addresses cross-view isolated sign language recognition, where models trained on frontal-view videos must classify signs recorded from left and right angles. It claims that a two-stage ensemble of Video Swin Transformers—Small, Base, and Large, combined first within each modality and then across RGB and depth—improves robustness to viewpoint change. On the MM-WLAuslan benchmark, the ensemble reaches 20.29% top-1 accuracy for RGB and 24.53% for RGB-D, outperforming every single backbone reported (best single: 17.51% on RGB). The solution ranked third in both tracks of the WWW 2025 CV-ISLR challenge. The load-bearing idea is that models of different capacities capture complementary granularities of gesture features, so averaging their predictions generalizes better to unseen viewpoints.","feed_headline":"Ensembling three Video Swin models boosts cross-view sign accuracy","feed_subtitle":"Two-stage RGB-depth ensemble lands 3rd at the WWW 2025 CV-ISLR challenge and beats every single backbone on RGB.","key_machinery":"The central object is the two-stage ensemble of Video Swin Transformer (VST) models, where VST is a hierarchical video transformer whose 3D shifted-window self-attention jointly models spatial and temporal features. The first stage performs single-modal classification ensemble: weighted sums of the Small, Base, and Large VST predictions within the RGB branch and within the depth branch, with weights $0.4$, $0.4$, $0.2$ respectively. The second stage performs multi-modal fusion ensemble: a weighted combination of the RGB and depth branch outputs with weights $0.65$ and $0.35$. The aggregation is a fixed weighted average of model outputs, not a learned fusion, which keeps the method simple and directly shows the effect of model-capacity diversity plus cross-modal complementarity.","core_discovery":"The central claim is that ensemble learning, applied at two levels, materially improves cross-view sign language recognition over any single Video Swin Transformer. Within each modality, the outputs of VST-Small, VST-Base, and VST-Large are combined with fixed weights $\\lambda_l=0.4$, $\\lambda_b=0.4$, $\\lambda_s=0.2$; for the RGB-D task, the RGB and depth branch outputs are then combined with weights $\\lambda_r=0.65$, $\\lambda_d=0.35$. The ensemble yields 20.29% top-1 accuracy on the RGB track and 24.53% on the RGB-D track, compared with 17.51% for the best single RGB model and 17.58% for the best single depth model. The paper attributes this gain to the complementary strengths of models with different dimensions in handling viewpoint variability and gesture complexity.","pith_inferences":["The authors leave implicit that the reported gain is sensitive to the chosen weights; if $\\lambda_l=0.4, \\lambda_b=0.4, \\lambda_s=0.2$ and $\\lambda_r=0.65, \\lambda_d=0.35$ were tuned on the validation view, the ensemble advantage could shrink on a truly novel viewpoint.","An extension the paper does not explore is replacing the fixed weights with a learned aggregation or a small gating network, which might preserve the robustness gain while adapting to view-specific reliability.","The two-stage ensemble idea could be applied to other modality pairs, such as RGB plus skeletal keypoints or RGB plus optical flow; if the cross-modal complementarity is the driver, similar gains should appear there.","Since the paper reports no single-model RGB-D baseline, a direct test of whether multi-modal fusion itself helps—versus just ensembling more models—would require training individual VST-S/B/L on RGB-D input and comparing their average to the reported 24.53%."],"forward_implications":["On the MM-WLAuslan benchmark, capacity diversity alone—without changing training data, loss, or architecture family—yields a 2.78-point top-1 gain on the RGB track over the best single VST model.","The same recipe transfers to depth input, where the single best model reaches 17.58% and the full RGB-D ensemble reaches 24.53%, indicating that depth and RGB carry complementary cues for cross-view recognition.","Because the method is a fixed-weight averaging of off-the-shelf VST backbones, it is a cheap robustness lever: any group with pretrained VST variants can reproduce it without redesigning the model.","The third-place finish in both tracks of the WWW 2025 challenge provides a reproducible baseline; the paper's code release lets future work compare directly against this ensemble.","The large gap to the top-ranked teams (56.87% RGB, 57.97% RGB-D) shows that the ensemble alone is not a complete solution to cross-view sign recognition, but it establishes a lower bound that specialized methods must beat."],"supporting_citations":[{"why":"Defines the Video Swin Transformer backbone that all single models and the ensemble are built on.","marker":"[13]"},{"why":"Introduces the MM-WLAuslan dataset and challenge setup; supplies the train/validation/test split and the reported accuracy numbers.","marker":"[20]"},{"why":"Provides the original Swin Transformer shifted-window mechanism that the video extension builds on.","marker":"[12]"},{"why":"Documents general benefits of ensemble deep learning, motivating the paper's core strategy.","marker":"[1]"},{"why":"Another ensemble-learning reference cited to justify aggregating multiple neural network predictions.","marker":"[43]"},{"why":"Applies ensemble-based transfer learning to Indian Sign Language recognition, showing precedent for ensembling in the ISLR domain.","marker":"[16]"}],"fun_headline_variants":["Three-model ensemble takes 3rd in cross-view sign language challenge","Ensemble of Video Swin models lifts cross-view sign recognition to 3rd","Two-level ensemble ranks 3rd in CV-ISLR challenge","Cross-view sign language: ensemble method beats single models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed improvement from ensembling rests on the specific hand-picked weights (0.4, 0.4, 0.2 for the single-modal stage and 0.65, 0.35 for the fusion stage) that the paper reports without any sensitivity analysis, so the gain over the best single model could be an artifact of weights tuned to the validation view rather than evidence of robust complementarity.","fun_headline_variants_meta":{"raw":{"variants":["Three-model ensemble takes 3rd in cross-view sign language challenge","Ensemble of Video Swin models lifts cross-view sign recognition to 3rd","Two-level ensemble ranks 3rd in CV-ISLR challenge","Cross-view sign language: ensemble method beats single models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001184,"raw_usage":{"total_tokens":4873,"prompt_tokens":913,"completion_tokens":3960,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":3885}},"tokens_in":529,"tokens_out":3960,"duration_ms":26626,"temperature":1.0,"reasoning_tokens":3885,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T12:57:46.770620+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-score or retrain the ensemble over a grid of weights—for example, $\\lambda_l$ and $\\lambda_b$ from 0 to 1 in steps of 0.1 with $\\lambda_s = 1-\\lambda_l-\\lambda_b$, and $\\lambda_r$ from 0 to 1 in steps of 0.1—and check whether the reported 20.29% RGB and 24.53% RGB-D top-1 accuracies lie on a broad plateau or a narrow peak; a narrow peak would indicate the ensemble gain is a weight-selection artifact, not a general property of capacity diversity.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents general benefits of ensemble deep learning, motivating the paper's core strategy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Another ensemble-learning reference cited to justify aggregating multiple neural network predictions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Applies ensemble-based transfer learning to Indian Sign Language recognition, showing precedent for ensembling in the ISLR domain."}],"review_version":1}