{"id":"6b65147e-d6b6-41f9-9b3c-66a1ec77da0b","arxiv_id":"2412.03498","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A Siamese bidirectional GRU network, fed six Mediapipe body-landmark frames per stride aligned with Procrustes analysis, reports 85 to 96 percent Rank-1 accuracy on CASIA-B, SZU, OU-MVLP, and Gait3D.","lead":"This paper describes a system that identifies people from the way they walk, using body-point data extracted from video and a neural network that compares walking sequences. The authors report over 86% accuracy on four standard gait datasets, which matters for surveillance and access control when faces are hidden or too far away.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation protocol for Rank-1 is unspecified on CASIA-B, SZU, and Gait3D; the missing gallery/probe split and matching rule makes the reported accuracies unverifiable.","rationale":"The reader's weakest_assumption focuses on whether the six-frame window spans a full gait cycle. That is a reasonable concern, but it is not the most load-bearing for the central claim. Even if the six frames do not cover a complete cycle, the reported accuracies could still be empirically correct on the evaluated sequences; the frame-selection deficiency would affect generalization claims but not necessarily invalidate the numbers. The evaluation protocol, by contrast, is the gatekeeper of the numbers themselves. The paper provides no test-time matching rule or gallery/probe definition for three of the four datasets, so the reported Rank-1 accuracies cannot be independently checked or reproduced. This is a more fundamental gap: a model with a perfect gait-cycle representation cannot be credited with an accuracy score if the scoring procedure is undefined. The reader's rationale does mention 'no protocol details for pair sampling' and 'no definition of D in contrastive loss,' but those points are listed alongside many other issues, and the formal weakest_assumption is the frame cycle. My load-bearing concern is distinct and, I argue, prior: without a protocol, the accuracy claims are not falsifiable. I therefore disagree with the reader's identification of the weakest assumption, while agreeing with the overall REJECT verdict. The concrete test—reproducing CASIA-B with a specified standard gallery/probe split—would settle whether the missing protocol hides a real problem or whether the reported numbers are robust. If the authors can provide the exact protocol and the numbers reproduce, the concern is resolved; if not, the abstract's accuracy claims require conditional qualification or rejection.","tokens_in":22311,"tokens_out":5048,"duration_ms":51506,"concrete_test":"Reproduce the CASIA-B result using the released GitHub code (github.com/promaprogga/Siamese-BiGRU-dualStack) with an explicitly specified standard protocol: train on the first 74 subjects, test on the remaining 50; for each test subject designate the first NM sequence as gallery and the remaining NM sequences as probes; extract the Siamese branch embedding (specified by the authors, e.g., the final hidden state or the post-dense output) and compute Rank-1 accuracy using Euclidean distance. If the mean over 54°, 90°, and 126° does not match the reported 95.7% (within a small tolerance), the claim is not reproducible under a standard protocol. The same check should be applied to SZU and Gait3D once the authors define their gallery/probe splits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the set of accuracy numbers in the abstract. These stand or fall on a well-defined, reproducible evaluation protocol. The paper only defines the protocol for OU-MVLP (Section 4.1: sequence '01' as gallery, '02' as probe). For CASIA-B it states only that 'the first 74 subjects are used for training, while the remaining subjects are reserved for testing' but never specifies which of the six NM sequences per test subject is the gallery, which are probes, or how per-view accuracies in Table II are computed. For SZU and Gait3D, no gallery/probe construction is given at all. The test-time use of the Siamese network is also unspecified: the paper describes training with contrastive loss (Eq. 11) and a Rank-1 metric (Eq. 12), but no evaluation-time distance matching rule—e.g., whether the hidden state of the last timestep, the concatenated bidirectional output, or the post-dense sigmoid output is used as the embedding, and whether ranking uses Euclidean or cosine distance. Without this, the numbers are uninterpretable; any of a wide range of ad hoc matching choices could produce different results. A further red flag appears in Section 4.6: Table I reports inter-subject Euclidean distances between encoded vectors as low as 0.002 (P1 vs P3), while Rank-1 accuracy is claimed to be 95.7%. Such a tiny inter-class margin is difficult to reconcile with near-perfect ranking unless the table is from a non-representative subset or the matching metric differs from the reported distance. This internal tension reinforces that the evaluation pipeline is not described precisely enough to support the headline results.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a gait recognition pipeline that extracts 33 Mediapipe body landmarks from video frames, selects six frames based on the left-foot landmark's x-coordinate, aligns the landmarks with Procrustes analysis, and feeds the resulting sequences into a Siamese network of two stacked bidirectional GRUs trained with a contrastive loss. The authors report Rank-1 accuracies of 95.7% on CASIA-B (NM condition), 94.44% on SZU RGB-D, 87.71% on OU-MVLP, and 86.6% on Gait3D, and compare these results against several appearance-based and model-based methods. The paper also includes an ablation study over RNN variants and a brief analysis of landmark subsets and clothing/bag effects.","tokens_in":22609,"tokens_out":4390,"duration_ms":45346,"significance":"If the reported accuracies were reproducible, the contribution would be noteworthy: a model-based method using only six frames of 2D landmarks would rival or exceed the state of the art on multiple benchmarks in both indoor and in-the-wild settings, with a lightweight architecture that could enable real-time gait identification. The use of Mediapipe landmarks and Procrustes alignment is a sensible design choice for view robustness. However, the significance of the paper is contingent entirely on the validity of the empirical evaluation; because the evaluation protocols are not fully specified and several reported numbers are internally inconsistent, the claimed advances cannot currently be assessed. The paper would make a stronger contribution if it provided a complete, reproducible evaluation protocol and reconciled the inconsistencies in the distance analysis.","major_comments":[{"comment":"The evaluation protocol is not defined for CASIA-B, SZU, or Gait3D. For CASIA-B, the paper only states that the first 74 subjects are used for training and the remaining for testing; it does not specify which of the six NM sequences per subject serve as gallery versus probe, how gallery and probe frames are sampled, or how the per-view accuracies at 54°, 90°, and 126° are computed. For SZU and Gait3D, no gallery/probe construction is provided at all. Additionally, the inference-time matching rule is missing: the paper does not state which model output (e.g., last hidden state, concatenated bidirectional output, or post-dense sigmoid value) is used as the embedding, nor which distance metric (Euclidean, cosine) is used for Rank-1 retrieval. Without these details, the central accuracy numbers in the abstract are unverifiable and unreproducible, so the paper's core claim does not meet standard evaluation requirements.","section":"Sections 4.1 and 4.7, Tables II-V"},{"comment":"The choice of 400 training pairs is based on observed performance: the text says 'As the number of pairs increases, the model’s performance improves. Based on this observation, we chose to utilize 400 pairs for training.' Figure 6 plots 'Dataset Pair Size and Accuracy' but does not specify whether this accuracy is on a validation set or the final test set used for the reported results. If the accuracy is a test-set measurement, then the model hyperparameters are tuned on the test set, which invalidates the reported generalization performance. Even if a validation set was used, the paper must say so explicitly and describe the validation protocol. The paper also omits how positive pairs are generated per subject (e.g., how many sequences per subject, whether the same sequence is compared with itself), how negative pairs are randomly sampled, and whether the 74 positive pairs for CASIA-B correspond to one pair per training subject or cross-sequence pairs, making the training procedure non-reproducible.","section":"Section 4.4 and Figure 6"},{"comment":"The Euclidean distances between encoded vectors of different individuals are implausibly small in light of the claimed 95.7% Rank-1 accuracy. For example, the distance between P1 and P3 is reported as 0.002, and the distance between P1 and P5 is 0.014. With such tiny inter-class margins, a nearest-neighbor or distance-based ranker would be expected to confuse these subjects almost always, unless the table is computed from a non-representative subset or the distance metric used for matching is different from the Euclidean distance shown here. The paper must clarify how these encoded vectors were produced, which network layer they correspond to, why such small distances still yield correct identification, and how Table I is consistent with the Rank-1 results in Table II. Without this reconciliation, the learned representation's discriminative quality is in question.","section":"Section 4.6, Table I"},{"comment":"The claim that six frames selected by monitoring the left foot landmark's x-coordinate 'progressing from the most negative to the most positive' constitute a complete gait cycle is not justified. In normal gait, the foot's x-coordinate oscillates with the swing and stance phases; a monotonic progression from negative to positive over six frames would typically span only a portion of a cycle (roughly a half-cycle of forward motion), not a complete cycle. The paper does not present evidence (e.g., a plot of the x-coordinate trajectory over time) or a reference that this selection yields a full cycle across viewpoints and subjects. This assumption is load-bearing because the fixed sequence length N=6 is the sole temporal input to the GRU; if the frames do not consistently align with a complete gait cycle, the model may be learning from incomplete or misaligned motion cues, which would affect the validity and transferability of the reported accuracies.","section":"Section 4.2"},{"comment":"The reported 86.6% Rank-1 accuracy on Gait3D is well above published results on that dataset (e.g., GaitBase 64.6% as cited by the authors), yet the paper does not describe how the model and preprocessing are adapted to Gait3D. Gait3D provides 3D SMPL models and multi-view videos, while the method described in Section 4.2 is based on Mediapipe 2D landmarks extracted from video frames. The text only says that 'In the Gait3D dataset, sequences are sourced from 4,000 subjects, with 3,000 subjects used for training and 1,000 for testing,' which is insufficient to understand whether landmarks are computed from RGB frames, projected from 3D SMPL, or obtained in some other way, and how the six-frame selection criterion is applied in an in-the-wild setting with varying camera viewpoints. The absence of these details makes the Gait3D result unverifiable and suggests a possible mismatch between the described pipeline and the dataset used for evaluation.","section":"Sections 4.2 and 4.4, Table V"}],"minor_comments":[{"comment":"The notation '*' is used for operations in the GRU equations but is never defined; depending on the context it could mean matrix multiplication, element-wise multiplication, or concatenation. This ambiguity makes the architectural description difficult to follow.","section":"Section 3, Equations (1)-(9)"},{"comment":"The contrastive loss is credited to reference [84] (Khosla et al., Supervised Contrastive Learning), but that reference describes a different loss formulation. The standard contrastive loss for Siamese networks typically cites Hadsell et al. (2006); the authors should cite the appropriate source and ensure the loss function is correctly referenced.","section":"Section 4.5, Equation (11)"},{"comment":"Figure 6's caption 'Relations Between Dataset Pair Size and Accuracy' does not specify which dataset is used, which accuracy metric is plotted, or whether the plotted points are train or test accuracies; the figure should be self-contained and clearly annotated.","section":"Section 4.4 and Figure 6"},{"comment":"The table lists subjects P1, P3, P4, P5, P8, P9, P12, P13, omitting subjects P2, P6, P7, P10, and P11 without explanation; if these subjects were excluded for a reason, the authors should state why.","section":"Section 4.6, Table I"},{"comment":"The paper contains numerous typographical and grammatical errors that should be corrected, including 'avaialble' in the abstract, 'procrustus' in Section 6, and the phrase 'two bidirectional bidirectional Gated Recurrent Units' in Section 3.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper reports impressive accuracy numbers, but the evaluation protocol is critically underspecified for three of the four datasets, the training pair count appears to be selected based on test-set performance, and the reported distance values in Table I are inconsistent with the claimed Rank-1 accuracy. These issues are central to the paper's contribution and would require a fundamentally more rigorous experimental setup to resolve. The Gait3D result in particular appears implausible given the described pipeline and the absence of dataset adaptation details. I recommend rejection on the grounds that the central claims are currently unverifiable and internally inconsistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the method is a standard Siamese biGRU with contrastive loss on Mediapipe landmarks; the more interesting piece is the preprocessing, using the left-foot landmark to define a six-frame window and Procrustes alignment to reduce view effects. The paper does run on four datasets, including Gait3D, and includes an ablation over RNN variants, so it is not a bare claim.\n\nWhat it does well: the preprocessing choice is sensible, the ablation is a real comparison, and the authors provide a GitHub link. The limitations discussion on loose clothing and bags is honest.\n\nThe trouble is the evaluation. Only OU-MVLP gets a full protocol: sequence '01' as gallery, '02' as probe. For CASIA-B, SZU, and Gait3D, there is no stated gallery/probe split, no description of how the Siamese embedding is produced at test time, and no matching rule (Euclidean? cosine? which hidden state?). The rank-1 numbers in the abstract therefore cannot be reproduced or checked from the paper. That is the load-bearing flaw, not a minor formatting issue.\n\nThere are additional concerns. The 400 training pairs were selected after looking at the accuracy curve in Fig. 6, which is post-hoc tuning of the evaluation set. No error bars are reported. CASIA-B is compared only on the NM condition. Table I shows inter-subject Euclidean distances as low as 0.002, which is hard to square with 95.7% rank-1 unless the matching distance is not the one displayed. And the Gait3D result, 86.6%, is 22 points above the best cited baseline; without a protocol, that gap is a red flag rather than a triumph.\n\nThe paper is not nonsense. The method is a reasonable engineering combination, and if the GitHub code is actually available, a patient reader could reverse-engineer the missing details. But as a scientific claim, the reported accuracies are unverified.\n\nWho this is for: practitioners who want a lightweight model-based gait recognizer and are willing to test it themselves. I would not cite the accuracy numbers. I would send it to a referee rather than desk-reject, because the missing protocol is fixable and the underlying approach is worth a look, but the referee should require the full evaluation protocol before any acceptance.","headline":"Plausible lightweight gait recognizer whose headline accuracies are not yet checkable because the evaluation protocol is mostly unspecified.","tokens_in":23191,"tokens_out":4745,"would_cite":false,"duration_ms":45743,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Siamese biGRU-dualStack network identifies people from six frames of 33 body landmarks, reporting 95.7% rank-1 accuracy on CASIA-B.","keywords":["gait recognition","body landmarks","pose estimation","Procrustes analysis","Siamese network","bidirectional GRU","cross-view gait recognition","biometric identification"],"falsifier":"Run the exact pipeline on CASIA-B with six frames selected uniformly at random instead of by the left-foot x-extremum rule, keeping everything else identical. If rank-1 accuracy at 90 degrees falls well below the reported 96.1%, the result is driven by the cycle-spanning frame selection; if it stays about the same, the sparse window itself, not the cycle heuristic, carries the identity signal.","tokens_in":22095,"feed_emoji":"🚶","tokens_out":4907,"duration_ms":44402,"temperature":0.7,"pith_summary":"The paper aims to show that a person's identity can be recognized from a very short window of body landmarks: six frames of 33 landmark points, aligned by Procrustes analysis and processed by a Siamese network of two stacked bidirectional GRUs. If the reported accuracies hold, gait becomes a practical biometric at a distance, requiring only a few pose estimates rather than full silhouettes or dense 3D body models. The authors report rank-1 accuracies of 95.7% on CASIA-B, 94.44% on SZU RGB-D, 87.71% on OU-MVLP, and 86.6% on Gait3D. The ablation study shows the dual-stack bidirectional GRU outperforms RNN, LSTM, GRU, and single-direction variants on both SZU and CASIA-B.","feed_headline":"Gait ID hits 95.7% from just six pose frames","feed_subtitle":"Sparse landmarks plus Procrustes alignment let a Siamese BiGRU beat prior models on four gait datasets.","key_machinery":"The load-bearing machinery is the combination of a sparse pose representation and a paired sequence encoder. Each person is represented as a 6 x 33 x 3 tensor of landmarks: six time steps, 33 landmark points, and x/y/z coordinates. Procrustes analysis, specifically generalized Procrustes alignment, removes rigid-body differences among subjects and viewpoints before the sequence reaches the network. The Siamese biGRU-dualStack architecture then encodes both forward and backward temporal context with two stacked bidirectional GRU layers per branch, concatenates the two branch outputs, passes them through a 1x1 dense layer and a sigmoid, and is trained with contrastive loss. The claim is that this compact representation captures identity-bearing gait dynamics without needing silhouettes or dense 3D models.","core_discovery":"The paper's central claim is that spatial alignment of sparse landmark sequences plus bidirectional recurrent encoding is enough for competitive gait recognition across indoor and in-the-wild datasets. Concretely, the pipeline extracts six frames per gait sample, selects them so the left-foot landmark's x-coordinate goes from most negative to most positive (intended to span a full walking cycle), collects 33 landmarks per frame, aligns the configurations with Procrustes analysis to remove translation, rotation, and scale differences across viewpoints, and feeds the resulting 594-value sequence into a Siamese network whose branches each contain two stacked bidirectional GRUs with 128 units. A contrastive loss trains the network to pull same-person pairs together and push different-person pairs apart. The authors report that this beats prior model-based and appearance-based methods on SZU, OU-MVLP, and Gait3D and is competitive on CASIA-B.","pith_inferences":["The paper implicitly treats the six-frame left-foot-extremum rule as a universal gait-cycle surrogate; a testable extension would compare this heuristic against ground-truth cycle segmentation across viewpoints and walking speeds.","The strong Gait3D result (86.6% versus 64.6% for the GaitBase baseline) suggests landmark sparsity plus alignment may transfer to unconstrained settings better than silhouette methods; ablating Procrustes alignment on Gait3D would reveal how much of that gain comes from alignment itself.","Because the input is only 594 values per sample, the model is plausibly fast enough for real-time edge deployment, but the paper does not report inference latency, so measuring it would be a natural next step.","The contrastive pairing strategy (400 pairs for CASIA-B and SZU, a 1:2 positive-negative ratio for Gait3D) is chosen without a sensitivity analysis; varying pair count and ratio could show whether performance depends on this choice."],"forward_implications":["Person identification can be performed from as few as six pose frames, so surveillance cameras with modest frame rates could support gait-based re-identification at a distance.","The approach generalizes across contrasting datasets (indoor, RGB-D, multi-view large population, and in-the-wild), suggesting the method is not tied to a single capture setup.","Landmark reduction from 33 to lower-body-only landmarks (23-32) costs little accuracy, so cheaper or simpler pose extraction may be feasible in practice.","The ablation implies that bidirectional, stacked recurrent encoding is the ingredient that matters most; simpler recurrent variants lose several points on both SZU and CASIA-B."],"supporting_citations":[{"why":"Supplies the pose-estimation landmark extractor that produces the 33 body landmarks used as the model's input.","marker":"[61]"},{"why":"Provides the six-frame sequence setup and the GaitNet baseline whose protocol the paper follows.","marker":"[39]"},{"why":"Supplies the Procrustes analysis method used to align landmark configurations across views.","marker":"[82]"},{"why":"Defines the contrastive loss used to train the Siamese network on same-person and different-person pairs.","marker":"[84]"},{"why":"Supplies the CASIA-B dataset used for the main cross-view evaluation.","marker":"[74]"},{"why":"Supplies the SZU RGB-D dataset used for a second indoor evaluation.","marker":"[78]"},{"why":"Supplies the OU-MVLP large-population dataset used for cross-view evaluation.","marker":"[80]"},{"why":"Supplies the Gait3D in-the-wild dataset used to test generalization beyond controlled indoor capture.","marker":"[81]"}],"fun_headline_variants":["Six pose frames push gait ID to 95.7% accuracy","Sparse landmarks + Siamese BiGRU: gait ID above 95%","Gait recognition at 95.7% using six aligned pose frames","Six pose landmarks yield 95.7% gait ID accuracy","Sparse landmarks: Siamese BiGRU tops gait recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the assumption that six frames chosen by watching the left-foot landmark's x-coordinate move from its most negative to its most positive value actually span a complete gait cycle, so the network sees the whole walking rhythm and not a partial or viewpoint-distorted fragment.","fun_headline_variants_meta":{"raw":{"variants":["Six pose frames push gait ID to 95.7% accuracy","Sparse landmarks + Siamese BiGRU: gait ID above 95%","Gait recognition at 95.7% using six aligned pose frames","Six pose landmarks yield 95.7% gait ID accuracy","Sparse landmarks: Siamese BiGRU tops gait recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000815,"raw_usage":{"total_tokens":3549,"prompt_tokens":901,"completion_tokens":2648,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":2555}},"tokens_in":517,"tokens_out":2648,"duration_ms":18942,"temperature":1.0,"reasoning_tokens":2555,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:19:57.452640+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact pipeline on CASIA-B with six frames selected uniformly at random instead of by the left-foot x-extremum rule, keeping everything else identical. If rank-1 accuracy at 90 degrees falls well below the reported 96.1%, the result is driven by the cycle-spanning frame selection; if it stays about the same, the sparse window itself, not the cycle heuristic, carries the identity signal.","supporting_citations":[{"cited_title":"Human pose estimation using mediapipe pose and optimization method based on a humanoid model,","cited_arxiv_id":null,"evidence_quote":"Supplies the pose-estimation landmark extractor that produces the 33 body landmarks used as the model's input."},{"cited_title":"Gaitnet: An end-to-end network for gait based human identification,","cited_arxiv_id":null,"evidence_quote":"Provides the six-frame sequence setup and the GaitNet baseline whose protocol the paper follows."},{"cited_title":"Procrustes methods in the statistical analysis of shape,","cited_arxiv_id":null,"evidence_quote":"Supplies the Procrustes analysis method used to align landmark configurations across views."},{"cited_title":"Supervised contrastive learning,","cited_arxiv_id":null,"evidence_quote":"Defines the contrastive loss used to train the Siamese network on same-person and different-person pairs."},{"cited_title":"A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the CASIA-B dataset used for the main cross-view evaluation."},{"cited_title":"A large rgb-d gait dataset and the baseline algorithm,","cited_arxiv_id":null,"evidence_quote":"Supplies the SZU RGB-D dataset used for a second indoor evaluation."},{"cited_title":"Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the OU-MVLP large-population dataset used for cross-view evaluation."},{"cited_title":"Gait recognition in the wild with dense 3d representations and a benchmark,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gait3D in-the-wild dataset used to test generalization beyond controlled indoor capture."}],"review_version":1}