{"id":"d9694ab6-8bbe-4c62-9588-d454b8da4644","arxiv_id":"2501.12086","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A graph neural network that learns separate joint-connection patterns for each feature channel and each time frame reports top or near-top accuracy on four skeleton gesture and action benchmarks.","lead":"Researchers built a new graph neural network, DSTSA-GCN, that recognizes hand gestures and body actions from 3D skeleton data by learning separate connection patterns for each feature channel and each moment in time. It reports top or near-top accuracy on four standard gesture and action benchmarks, though without released code or uncertainty estimates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SHREC'17 headline numbers are likely inflated because the test split was used to select K, theta, STCA, MS-TCN branches, and initialization; the SOTA claim rests on these numbers.","rationale":"The reader's weakest assumption identified the same load-bearing concern: SHREC'17 test-set-driven hyperparameter selection. This is the most direct threat to the central claim because the reported SHREC'17 numbers are the headline gesture-recognition results and the only benchmark on which the model is claimed to be state-of-the-art in both gesture settings. The ablation tables (Tables 2–5) are all conducted on SHREC'17, and the final configuration is selected by test accuracy, so the reported 97.74% and 95.37% are likely optimistic. Other issues noted by the reader—such as the parameter-count claim being inconsistent with the equations, and several second/third-best results on NTU benchmarks—are real but secondary; they weaken the 'state-of-the-art' phrasing but do not directly invalidate the empirical support as much as the selection bias does. The concrete test proposed here would quantify the inflation and determine whether the central claim survives. I agree with the reader's conditional verdict: the paper should be accepted only if code is released, multiple runs are reported, and SHREC'17 is re-evaluated with a validation split. No change to the reader's verdict is needed.","tokens_in":18648,"tokens_out":5924,"duration_ms":59995,"concrete_test":"Re-run the SHREC'17 experiment with a proper validation split. Hold out, for example, 200 of the 1960 training sequences as a validation set. Perform the same grid search over Tables 2–5 on this validation set, select the best configuration, then retrain on all 1960 training sequences with that configuration and evaluate on the official 840-sequence test set exactly once. If the resulting 14/28-gesture accuracies are within 0.5% of 97.74/95.37, the selection bias is likely negligible; if they drop by more than 1%, the reported numbers are optimistic and the state-of-the-art claim on SHREC'17 is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim of state-of-the-art performance on SHREC'17 rests on the 97.74% and 95.37% fused accuracies in Table 7. However, the exhaustive ablation study is conducted entirely on the SHREC'17 dataset (Section 4.3), and Tables 2–5 select the number of groups K, the activation theta, the STCA construction, the MS-TCN branch set, and the static-topology initialization by comparing test-set accuracies. SHREC'17 provides a fixed test split of 840 sequences, and the paper reports no held-out validation set for these choices. Using the test split for model selection makes the reported accuracies optimistic estimates of generalization; the true held-out performance is unknown and likely lower. This is not merely a theoretical concern: the selected configuration (e.g., K=8, Tanh, STCA, branches H in Table 4) was chosen precisely because it maximized test accuracy, so comparisons with baselines that did not tune on this test set are unfair. If the inflation exceeds about 1%, DSTSA-GCN would fall below STDA-GCN and possibly TD-GCN on the 28-gesture benchmark, directly undermining the claimed state-of-the-art status. The same test split is also used to motivate the design narrative, so the empirical support for the method's components is weaker than presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents DSTSA-GCN, a graph convolutional network for skeleton-based gesture and action recognition. The architecture combines a grouped channel-wise graph convolution (GC-GC) and a grouped temporal-wise graph convolution (GT-GC) with a multi-scale temporal convolution (MS-TCN), and introduces a Spatio-Temporal Coordinate-Aware (STCA) feature transformation. The authors report fused accuracies of 97.74% and 95.37% on SHREC'17 Track, 95.04% and 93.57% on DHG-14/28, 92.78% and 97.03% on NTU-60, and 89.12% and 90.97% on NTU-120, and claim state-of-the-art performance on these benchmarks.","tokens_in":19038,"tokens_out":5524,"duration_ms":51854,"significance":"The proposed modules are clearly motivated, and the ablation study (Tables 1-5) is thorough in exploring each design choice. The per-subject DHG results in Table 6 and the parameter/FLOP reporting are useful. If the reported accuracies are unbiased, the method is competitive with recent GCN baselines at similar parameter counts. However, the headline SHREC'17 result is compromised by test-set-based model selection, and the state-of-the-art claim is not supported by the authors' own tables for several benchmarks. The method's technical contribution is plausible, but the empirical evidence needs strengthening.","major_comments":[{"comment":"The SHREC'17 test split is used to select the number of groups K, the activation θ, the STCA construction, the MS-TCN branch configuration, and the static-topology initialization (Tables 2-5), and the same split is then used to report the final accuracies in Table 7 (97.74% and 95.37%). Since no validation split is described in Section 4.2, these hyperparameters are effectively fitted to the test set, making the SHREC'17 comparison with baselines that did not tune on this split optimistic. The paper should either introduce a validation-based selection procedure or explicitly quantify the expected optimism, for example by reporting results for a fixed default configuration.","section":"4.3, Tables 2-5"},{"comment":"The abstract states that DSTSA-GCN achieves state-of-the-art performance on SHREC'17, DHG-14/28, NTU-RGB+D, and NTU-RGB+D-120, but the paper's own tables contradict this: Table 8 shows that on NTU-60 X-Sub DSTSA-GCN (92.78%) is below BlockGCN and LG-SGNet (93.1%), and on NTU-120 X-Sub and X-Setup the method is below BlockGCN (89.12 vs 90.3 and 90.97 vs 91.5), while Table 7 shows the method is second-best on SHREC'17 28-gesture (95.37 vs 95.84 for STDA-GCN). The performance claims should be revised to match the reported numbers.","section":"Abstract; Section 4.4, Tables 7-8"},{"comment":"Section 3.3 states that 'the number of parameters in our approach is independent of the number of groups,' but the static topology tensor A ∈ R^{K×V×V} in Eq. 22 introduces K·V^2 parameters, which scales linearly with K. This is also inconsistent with Table 2, where the parameter counts change with K (1.96M, 1.97M, 1.99M, 1.97M). The statement should be corrected or clarified to specify which parameter subset is group-independent.","section":"3.3, Eq. 22"},{"comment":"No error bars or multiple-seed results are reported for any dataset. The SHREC'17 and NTU results are single-split evaluations, and the DHG result is a mean over 20 leave-one-subject-out folds without a standard deviation. Since many of the reported improvements over baselines are between 0.1 and 0.5 percentage points, the statistical significance of the claimed gains is not established. At minimum, the authors should report standard deviations or multiple runs.","section":"4.2; Tables 7-8"}],"minor_comments":[{"comment":"Equation 20 is missing the multiplication sign between the feature and adjacency matrices; as typeset it is ambiguous.","section":"Eq. 20"},{"comment":"Equation 14 (TGP) does not specify over which dimension the softmax is taken; the summation over t and the softmax over a scalar expression need clarification.","section":"Eq. 14"},{"comment":"Section 2.2 repeats the citations [10, 11, 12, 13, 14] twice in one sentence; the duplicate citation list should be removed.","section":"2.2"},{"comment":"The code link in the abstract is given as a URL without a release date; please confirm the repository is public at the time of publication.","section":"Abstract"},{"comment":"In Table 7, single-modality DSTSA-GCN results are reported for SHREC'17 but not for DHG-14/28; including them would help isolate the contribution of the fusion.","section":"Table 7"}],"recommendation":"major_revision","confidential_remarks":"The test-set tuning issue is the main barrier to acceptance; if the authors can re-run selection on a validation split or clearly frame the SHREC'17 numbers as best-on-test and temper the SOTA claim, the paper could be publishable. The parameter-count contradiction is minor but should be fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent incremental GCN paper for skeleton-based gesture and action recognition, and the parallel combination of channel-specific and frame-specific non-shared topologies under one shared STCA transform is genuinely new. But the SOTA claim on SHREC'17 is built on test-set tuning, and the margin over the closest temporal-topology baseline is a fraction of a single test sequence. Treat the rankings as provisional.\n\nWhat's actually new: the ingredients are mostly borrowed and properly cited—channel-wise dynamic topologies from CTR-GCN, temporal-wise topology from TD-GCN and DG-STGCN, grouped convolution from the usual suspects. The new part is running GC-GC and GT-GC in parallel with a shared spatio-temporal coordinate-aware transform and adding the grouped multi-scale MS-TCN. That combination does not appear in the cited literature. The paper also does several things well: the ablations are systematic and cover each component; the findings are consistent (GC-GC is the workhorse, GT-GC gives a real but smaller bump, MS-TCN cuts FLOPs by over a third while improving accuracy); and the second- or third-best results on several benchmarks are reported plainly rather than hidden.\n\nSoft spots, in order of weight. First: the SHREC'17 headline numbers are inflated by construction. Tables 2–5 select K, the activation, the STCA layout, the MS-TCN branches, and the initialization by trying variants on the SHREC'17 test split, and the same split then appears as the headline result. On the 28-class task the margin over TD-GCN is 0.01%—a fraction of a single test sequence—so even a small optimism bias flips the ranking, and the paper's own table shows STDA-GCN ahead on that benchmark. The DHG and NTU numbers are more believable: DHG uses a leave-one-subject-out protocol and the gains there are larger (1.14 and 2.17 points over TD-GCN); the NTU results show no sign of test-set tuning. Second, no error bars or multiple runs, so all the small margins are unquantified. Third, the Section 3.3 claim that parameter count is independent of the number of groups is contradicted by Table 2 (1.96M at K=3, 1.99M at K=8)—minor, since the variation is tiny, but sloppy. Fourth, code is promised, not shipped, and the IB-loss experiment is mentioned but never shown.\n\nWho it's for: people working on GCN-based skeleton recognition who want to see the next incremental step past CTR-GCN and TD-GCN. It deserves a serious referee, not a desk reject. I'd send it to review with a clear request: rerun SHREC'17 with a proper validation split, report multiple runs and variance, correct the parameter-count statement, and either ship the code or soften the efficiency claims.","headline":"A competent incremental GCN paper whose parallel channel-wise/temporal-wise topology design is genuinely new, but whose SHREC'17 SOTA claim rests on test-set tuning and a margin of a fraction of one test sequence.","tokens_in":19492,"tokens_out":12600,"would_cite":true,"duration_ms":109184,"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":"DSTSA-GCN claims that per-channel and per-frame topology graphs, plus multiscale temporal convolution, push skeleton gesture recognition to 97.74% on SHREC'17 and 95.04% on DHG.","keywords":["skeleton-based gesture recognition","graph convolutional networks","dynamic topology modeling","channel-wise graph convolution","temporal-wise graph convolution","multi-scale temporal convolution","human action recognition","spatio-temporal modeling"],"falsifier":"Retrain DSTSA-GCN on SHREC'17 with all design choices (K, theta, STCA variant, MS-TCN branch set, static initializer) selected by cross-validation on the training split alone, then evaluate once on the fixed test split; if accuracy falls materially below 97.74% and 95.37%, the reported gain is at least partly an artifact of test-set selection.","tokens_in":18428,"feed_emoji":"🤟","tokens_out":7082,"duration_ms":63405,"temperature":0.7,"pith_summary":"The paper is trying to establish that skeleton-based gesture recognition improves when graph convolution does not force one adjacency pattern across all channels and all frames. The proposed DSTSA-GCN builds separate dynamic topologies for channel groups and for temporal positions, and it pairs these with a multi-scale temporal convolution. The authors report that this design reaches top accuracies on SHREC'17 Track (97.74% for 14 gestures, 95.37% for 28 gestures) and DHG-14/28 (95.04% and 93.57%), and competitive results on NTU RGB+D and NTU RGB+D 120. A careful reader would care because the grouped construction keeps the parameter count roughly independent of the number of groups, so the added expressiveness does not come from a large increase in model size.","feed_headline":"A graph that shifts per frame lifts gesture accuracy to 97.74%","feed_subtitle":"New design models channel- and frame-specific topology, beating prior graph networks on hand and full-body benchmarks.","key_machinery":"The load-bearing object is the pair of grouped topology-modeling operations, GC-GC and GT-GC, together with MS-TCN. For a channel group, the adjacency matrix is $\\hat{A}_c = \\alpha A_c^d(X) + A_g$, where $A_c^d(X)$ is a sample-dependent distance matrix computed from transformed joint features, $\\alpha$ is a trainable scalar, and $A_g$ is a static graph shared within the group. GT-GC applies the same formula after pooling over channels, producing frame-specific graphs. The grouping trick keeps the number of static graphs equal to the number of groups $K$ rather than the number of channels, so parameter count stays nearly flat while each channel still gets a distinct dynamic topology. MS-TCN contributes multiscale temporal receptive fields through parallel dilated temporal convolutions.","core_discovery":"The central claim is that spatial and temporal topology should both be dynamic and non-shared, and that grouping lets the model capture multiscale joint interactions without inflating parameters. In the Group Channel-wise Graph Convolution, each channel receives its own adjacency matrix built from a learnable static graph plus a sample-dependent graph computed from distances between transformed joint features; the Group Temporal-wise Graph Convolution applies the same construction along the temporal axis, so different frames can use different interaction graphs. A Spatio-Temporal Coordinate-Aware module injects position information during feature transformation, and the Multi-Scale Temporal Convolution replaces the single temporal kernel with parallel branches at dilations {1,2,3,4} plus a max-pooling branch and a shortcut. Ablations on SHREC'17 attribute the largest single gain to the channel-wise module, and fused four-modality results on gesture and action benchmarks exceed the results of CTR-GCN and TD-GCN in most reported settings.","pith_inferences":["The paper does not test whether its hyperparameters were chosen on a validation split rather than the SHREC'17 test split; if a validation-only retraining reproduces the gains, the same grouped topology idea likely transfers to other graph time-series where connectivity changes over time, such as human-object interaction or multi-agent motion.","The frame-specific topology learned by GT-GC could serve as an interpretability tool: class-activation maps and temporal topology visualizations suggest that particular frames carry distinct joint interactions, and one could quantify how much each frame's graph contributes to the prediction.","Because the static graph is randomly initialized and learns dataset-level topology, the grouping strategy may be a general remedy for the local-bias problem of deep GCNs beyond skeleton data, applicable whenever a fixed adjacency matrix overly weights near neighbors."],"forward_implications":["On SHREC'17 Track, fusing joint, bone, joint-motion, and bone-motion modalities gives 97.74% for 14 gestures and 95.37% for 28 gestures, above the closest prior GCN results reported in Table 7.","On DHG-14/28, the same fusion reaches 95.04% and 93.57%, improvements of 1.94 and 3.07 points over CTR-GCN.","On NTU RGB+D, the model reaches 92.78% (Cross-Subject) and 97.03% (Cross-View); on NTU RGB+D 120 it reaches 89.12% (Cross-Subject) and 90.97% (Cross-Setup), matching or exceeding several recent baselines.","The grouped static graphs and MS-TCN lower parameter and FLOP counts relative to a standard TCN while improving accuracy, so the gain is not a parameter-count effect.","Ablations show each component contributes: removing GC-GC costs the most accuracy, and the combined STCA module outperforms separate spatial-only or temporal-only coordinate-aware modules."],"supporting_citations":[{"why":"Supplies the ST-GCN baseline whose fixed, partition-based adjacency matrix the paper extends.","marker":"[8]"},{"why":"Establishes the dynamic-graph line that GC-GC refines with per-channel topologies.","marker":"[10]"},{"why":"CTR-GCN is the channel-wise topology baseline the paper improves on, and its group design is the direct comparison.","marker":"[16]"},{"why":"Provides the dynamic spatial-temporal modeling motivation for making temporal topology non-shared.","marker":"[17]"},{"why":"Defines the DHG-14/28 dataset and leave-one-subject-out protocol used for gesture benchmarks.","marker":"[29]"},{"why":"TD-GCN is the closest prior temporal-decoupling competitor for gesture recognition and a direct comparison in Tables 7 and 8.","marker":"[34]"},{"why":"Defines the SHREC'17 Track dataset, its 14/28 class labels, and the training/test split used in ablations and final results.","marker":"[41]"},{"why":"Supplies the NTU RGB+D dataset and its Cross-Subject and Cross-View evaluation protocols.","marker":"[42]"},{"why":"Supplies the NTU RGB+D 120 dataset and its Cross-Subject and Cross-Setup evaluation protocols.","marker":"[43]"}],"fun_headline_variants":["Graphs that shift per channel and frame top gesture benchmarks","Adaptive topology modeling lifts skeleton recognition accuracy","DSTSA-GCN: dynamic space-time graphs win gesture benchmarks","Multiscale graph convolution surpasses prior skeleton gesture nets","Per-frame and per-channel graph shifts improve gesture AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the SHREC'17 test split was not used to select the number of groups, the activation function, the STCA design, the MS-TCN branches, and the static-topology initialization; if those choices were tuned on the test set, the reported 97.74% and 95.37% accuracies are optimistic.","fun_headline_variants_meta":{"raw":{"variants":["Graphs that shift per channel and frame top gesture benchmarks","Adaptive topology modeling lifts skeleton recognition accuracy","DSTSA-GCN: dynamic space-time graphs win gesture benchmarks","Multiscale graph convolution surpasses prior skeleton gesture nets","Per-frame and per-channel graph shifts improve gesture AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000949,"raw_usage":{"total_tokens":4083,"prompt_tokens":1011,"completion_tokens":3072,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":2994}},"tokens_in":627,"tokens_out":3072,"duration_ms":24258,"temperature":1.0,"reasoning_tokens":2994,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:31:38.369659+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain DSTSA-GCN on SHREC'17 with all design choices (K, theta, STCA variant, MS-TCN branch set, static initializer) selected by cross-validation on the training split alone, then evaluate once on the fixed test split; if accuracy falls materially below 97.74% and 95.37%, the reported gain is at least partly an artifact of test-set selection.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ST-GCN baseline whose fixed, partition-based adjacency matrix the paper extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the dynamic-graph line that GC-GC refines with per-channel topologies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CTR-GCN is the channel-wise topology baseline the paper improves on, and its group design is the direct comparison."},{"cited_title":"De Smedt, H","cited_arxiv_id":null,"evidence_quote":"Defines the DHG-14/28 dataset and leave-one-subject-out protocol used for gesture benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TD-GCN is the closest prior temporal-decoupling competitor for gesture recognition and a direct comparison in Tables 7 and 8."},{"cited_title":"De Smedt, H","cited_arxiv_id":null,"evidence_quote":"Defines the SHREC'17 Track dataset, its 14/28 class labels, and the training/test split used in ablations and final results."},{"cited_title":"Shahroudy, J","cited_arxiv_id":null,"evidence_quote":"Supplies the NTU RGB+D dataset and its Cross-Subject and Cross-View evaluation protocols."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NTU RGB+D 120 dataset and its Cross-Subject and Cross-Setup evaluation protocols."}],"review_version":1}