{"id":"4c137561-d91c-4aee-97ff-d2624afba5d5","arxiv_id":"2501.16608","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"GaitDCCR combines dynamic clustering parameters, density-weighted centroids, confidence-based pseudo-label refinement, and a contrastive teacher module to improve unsupervised gait recognition.","lead":"This paper introduces GaitDCCR, an unsupervised method that improves gait recognition by dynamically tuning clustering, weighting cluster centroids, and refining noisy pseudo-labels with confidence scores and a contrastive teacher network. With a GaitGL backbone it reaches 87.4% mean rank-1 accuracy on CASIA-B, surpassing prior unsupervised adaptation methods and some fully supervised models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 87.4% CASIA-B result likely depends on fixing the number of clusters C to the known 74-subject target training split, but the paper never states how C is chosen, weakening the 'unsupervised' claim.","rationale":"The reader's weakest assumption identifies precisely the cluster-count issue, and I agree that this is the single most load-bearing concern. The central result is a number produced by a clustering-based UDA pipeline, and every component — memory bank, DWC, CPR, CTM, and the final loss — is parameterized by C. If C is set to the known number of identities in the target training split, the method is not fully label-free in the sense claimed by the table heading 'without any labels on CASIA-B'. The dynamic epsilon schedule cannot compensate for a wrong cluster count, because epsilon only changes how tightly clusters are drawn, not how many clusters exist. The paper's silence on this point is not a minor implementation detail: it determines whether the unsupervised claim is meaningful. Other issues noted by the reader — under-specified loss in Eq. 13, the GaitSet/GaitGL inconsistency between Sec. 5.2.1 and Table 4, absent code, and lack of error bars — are real but secondary. They would affect confidence in the exact numbers, whereas the cluster-count issue affects the interpretation of the whole evaluation. The proposed check is straightforward and decisive: re-run with varied C and with an identity-count-free clustering method. This is an addressable concern rather than a fatal flaw, so the conditional verdict remains appropriate; no verdict change is needed.","tokens_in":17578,"tokens_out":6228,"duration_ms":63644,"concrete_test":"Run a controlled sweep over C on the OUMVLP→CASIA-B GaitGL setting: C ∈ {50, 74, 100, 150}, keeping all other hyperparameters fixed, and report mean rank-1, NM, BG, and CL for each. Also run a variant that does not fix C (e.g., HDBSCAN without a target-identity-count prior) and compare against the 87.4% headline. If the result at C=74 is substantially higher than neighboring values, the headline depends on leaked target-set identity count; if the paper's released code is available, additionally check whether C is hard-coded to len(target_train_identities).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GaitDCCR achieves 87.4% mean rank-1 on CASIA-B with GaitGL in a fully unsupervised domain adaptation setting, surpassing prior SOTA by 8.2%. The entire pipeline is indexed by the cluster count C: pseudo-labels are defined as yi ∈ {1,...,C} (Sec. 3), the memory bank has C centroids, the DWC density computation and CPR confidence matrix F are R^{B×C} (Eq. 8), the CTM latent set is drawn from C clusters, and the final loss sums over C clusters. Nowhere in Sec. 4 or Sec. 5 is the value or selection procedure for C stated. The only numeric identity counts given are the dataset split sizes: 74 training subjects for CASIA-B and 5,153 for OUMVLP. If the implementation sets C to the number of identities in the target training split, the method implicitly uses knowledge that a genuinely unlabeled deployment setting would not have. The dynamic epsilon schedule (Eq. 4) changes only cluster tightness; it cannot correct an incorrect cluster count. Moreover, the clustering algorithm itself is never named, only a probability threshold epsilon, so the role of C cannot be inferred from the text. If C is not derived from the known split, the paper still omits the mechanism by which C is obtained, making the reported result irreproducible and the 'without any labels' claim unverifiable. This is the most load-bearing assumption because every reported improvement flows through the clustering stage, and the headline comparison to SOTA depends on the validity of that stage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes GaitDCCR, an unsupervised domain adaptation method for gait recognition. The method has two stages: a clustering stage with Dynamic Clustering Parameters (DCP) and Dynamic Weighted Centroids (DWC), and a training stage with Confidence-based Pseudo-label Refinement (CPR) and a Contrastive Teacher Module (CTM). The authors evaluate on CASIA-B and OUMVLP using GaitSet and GaitGL backbones, reporting strong rank-1 accuracies, including 87.4% mean rank-1 on CASIA-B with GaitGL, which they claim surpasses prior state-of-the-art unsupervised methods and some fully supervised methods.","tokens_in":17941,"tokens_out":4473,"duration_ms":42907,"significance":"If the reported results hold, GaitDCCR would be a substantial empirical advance for unsupervised gait recognition, particularly for the challenging clothing-change scenario. The paper includes component-level ablations, an analysis of dynamic thresholds, and promises code release, which are strengths. However, the central claim is weakened by the undisclosed cluster count C, an internal contradiction between the text and Table 4, and an inconsistency between the described InfoNCE loss and the actual cross-entropy formulation. These issues need to be resolved before the results can be considered reproducible or the unsupervised claim fully verified.","major_comments":[{"comment":"The number of clusters C is never defined or reported anywhere, yet it is load-bearing for the entire pipeline: pseudo-labels are defined as yi ∈ {1,...,C}, the memory bank contains C centroids, F has dimensions R^{B×C} in Eq. (8), Eq. (10) assigns probabilities over C clusters, and Eq. (13) sums the loss over C clusters. The paper must state how C is chosen. If C is set to the number of known identities in the target training split (74 for CASIA-B and 5,153 for OUMVLP), that relies on information a genuinely unlabeled deployment setting would not have; this would need to be acknowledged and ablated. If C is estimated by another mechanism, that mechanism must be described. The dynamic epsilon schedule in Eq. (4) only changes cluster tightness and cannot correct an incorrect cluster count, so the omission is not benign.","section":"Section 3, Section 4.2, Algorithm 1"},{"comment":"The text states that \"using the GaitGL backbone consistently yields better accuracy compared to GaitSet in all tested conditions.\" This is directly contradicted by Table 4 for GREW→OUMVLP, where GaitSet reaches 68.5% mean rank-1 and GaitGL reaches 60.6%. The claim should be corrected or qualified, and the discrepancy between the two backbones in the GREW source setting should be discussed.","section":"Section 5.2.1 and Table 4"},{"comment":"The method is described as using InfoNCE loss, and Algorithm 1 line 18 says \"Computing InfoNCE loss,\" but the actual loss in Eq. (13) is a cross-entropy loss lce(Mq fi, y^s_i) applied to soft pseudo-labels. Section 3 Eq. (2) defines a separate InfoNCE loss that is not used in the final training objective. The authors should clarify which loss is actually optimized and reconcile the notation, since the distinction matters for reproducibility.","section":"Sections 3 and 4.6"},{"comment":"The text says \"samples with smaller density distance are more likely to be real samples, while samples with larger density distance are likely to be noise samples,\" but the defined quantity Di = σ(−(1/|Ci|) Σ d(i,j)) is larger when the average distance to other samples in the cluster is smaller. Thus, a \"larger density distance\" as defined corresponds to a sample closer to the cluster center, not a noisier sample. This makes the weighting direction in Eq. (6) ambiguous and should be corrected or clarified.","section":"Section 4.3, Eqs. (5)-(7)"},{"comment":"The ablation for the contrastive teacher module with and without data augmentation is unclear. Table 5 has columns for DCP, DWC, CPR, CTM, and CTM (w/o DA), but rows 7 and 8 both show four checkmarks and no visible distinction between the configurations. The text refers to experiments \"7\" and \"8\" for the DA ablation, but a reader cannot tell which row uses DA and which does not. Please fix the table formatting and explicitly annotate each row.","section":"Section 5.3.2 and Table 5"}],"minor_comments":[{"comment":"The subsection title says \"Effectiveness of CPR and CTM in the clustering phase,\" but the components described and ablated are in the training phase; please correct the wording.","section":"Section 5.3.2"},{"comment":"Equation (10) has an unclear denominator, \"d(˜xi,C j)P fCk i\", which appears to be a formatting or notation error. Please rewrite the normalization explicitly.","section":"Section 4.5.2, Eq. (10)"},{"comment":"Algorithm 1 line 17 uses y^k_i in the soft-label update, whereas Eq. (11) defines ys_i = β·Pi + (1−β)·y^i. The notation should be made consistent.","section":"Algorithm 1 and Eq. (11)"},{"comment":"The manuscript header on page 2 reads \"Preprint submitted to Nuclear Physics B,\" which appears to be an incorrect journal name for this computer vision paper; please update it.","section":"General"},{"comment":"The caption in Table 1 says \"CAISA-B\" instead of \"CASIA-B\"; please fix the typo.","section":"Table 1 caption"},{"comment":"Please cite the actual publication venues for GOUDA, UGRSF, and STANet in the comparison tables rather than only arXiv/venue names in the table cells, and double-check the reference formatting for these entries.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: if the numbers hold, this is a real advance in unsupervised gait recognition. The CASIA-B CL numbers with GaitGL are strong, and the ablations show each component pulling its weight. But the paper never says how the cluster count C is chosen, and that omission is load-bearing for the 'unsupervised' claim.\n\nWhat's new: the pipeline combines dynamic clustering threshold decay (DCP), density-weighted centroids (DWC), confidence-based soft pseudo-label refinement (CPR), and a contrastive teacher with clothing augmentation (CTM). Each piece borrows from existing ReID/contrastive work, but the combination for gait UDA is new, and the ablations demonstrate that DCP, DWC, CPR, and CTM each contribute. The CL improvement on CASIA-B (74.2 with GaitGL, versus 65.6 for UGRSF) is the headline result and is credible.\n\nSoft spots, in order of importance. First, C, the number of clusters, never appears in the implementation details. Equation (8), the memory bank, and the loss all depend on C, and the only identity counts given are the training split sizes (74 on CASIA-B, 5,153 on OUMVLP). If the code sets C to those numbers, the method is using a form of supervision that a genuinely unlabeled deployment would not have. The dynamic epsilon schedule adjusts cluster tightness, not the number of clusters, so it cannot fix a wrong C. The paper needs to state C and discuss how it would be estimated in practice. That is a genuine reproducibility gap, not a nit.\n\nSecond, the text claims GaitGL 'consistently yields better accuracy' than GaitSet, but Table 4 shows the opposite on GREW→OUMVLP (GaitSet 68.5, GaitGL 60.6). Minor internal contradiction, but it is their own data.\n\nThird, Eq. (13) is labeled InfoNCE but is soft cross-entropy over memory-bank logits; the relationship to Eq. (2) is unexplained. Minor.\n\nAlso: no error bars, and hyperparameters are tuned on the evaluation benchmarks. That is standard for this subfield, but worth saying out loud.\n\nWho this is for: researchers working on unsupervised gait recognition or clustering-based UDA for fine-grained biometrics. The method section is readable and the ablations are informative. I would send it to review—the cluster-count issue is addressable, and the experimental core deserves referee time. I would want the code released before final acceptance.","headline":"Strong empirical pipeline for unsupervised gait recognition, but the unstated cluster count C undermines the fully-unsupervised framing until it is disclosed.","tokens_in":18476,"tokens_out":2999,"would_cite":false,"duration_ms":28910,"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":"GaitDCCR establishes that unsupervised gait recognition can surpass fully supervised methods, reporting 87.4% mean rank-1 on CASIA-B.","keywords":["gait recognition","unsupervised domain adaptation","pseudo-label noise","dynamic clustering","contrastive learning","teacher-student training","silhouette data augmentation","soft pseudo-labels"],"falsifier":"Run GaitDCCR on a target set whose true identity count is withheld, estimating C from the feature distribution by silhouette analysis or a density-based cluster count, and report rank-1; if accuracy drops sharply relative to the run that sets C to the known training-split identity count, the central claim of fully label-free adaptation is falsified.","tokens_in":17341,"feed_emoji":"🚶","tokens_out":6625,"duration_ms":60956,"temperature":0.7,"pith_summary":"This paper tries to establish that unsupervised gait recognition can become nearly as accurate as supervised recognition if the noise in clustering-generated pseudo-labels is treated explicitly. It proposes GaitDCCR, which combines dynamic clustering thresholds, density-weighted cluster centroids, confidence-based soft pseudo-labels, and a contrastive teacher branch with silhouette augmentation. On CASIA-B with a GaitGL backbone it reports 87.4% mean rank-1, 8.2 points above the prior unsupervised state of the art, and above some fully supervised baselines. If this holds, label-free gait identification from silhouette sequences becomes practical for surveillance and security applications.","feed_headline":"Unsupervised gait recognition scores 87.4% rank-1, beating SOTA by 8.2","feed_subtitle":"Dynamic clustering thresholds and soft pseudo-labels close the gap with fully supervised gait recognition on clothing-change data.","key_machinery":"The method's central object is a soft pseudo-label pipeline built on a memory bank of cluster centroids. In the clustering stage, an exponential decay rule $\\epsilon = \\epsilon_0 \\eta^{\\text{epochs}}$ adjusts the cluster acceptance threshold, and each centroid is recomputed as a density-weighted sum $m'_i = \\sum_{x_i\\in C_i} w_i f_i$, where $w_i$ weights central samples more than edge samples. In training, a confidence matrix $F_{i,j} = p_{i,j}/\\sum_j p_{i,j}$ with $p_{i,j}=\\sigma(-d(i,j))$ converts hard labels into soft labels $\\hat{y}_i = \\alpha y_i + (1-\\alpha)F_i$, and a teacher branch, updated by exponential moving average, applies dilation and erosion augmentation to the body region, finds the $k$ nearest centroids, and produces the final soft label $y_i^s = \\beta P_i + (1-\\beta)\\hat{y}_i$; the student is trained with an InfoNCE loss against the memory bank. This machinery carries the argument because every reported gain is attributed to increasingly reliable pseudo-labels and centroids.","core_discovery":"The central claim is that pseudo-label noise caused by cross-domain clothing and viewpoint changes is the dominant failure mode in unsupervised gait recognition, and that it can be suppressed by a four-part mechanism: an exponentially decaying clustering threshold, density-weighted centroids, confidence-based refinement of one-hot pseudo-labels into soft labels, and a teacher-student contrastive module that uses body-region silhouette augmentation to find the true clusters of clothing-augmented samples. The paper reports that this mechanism lifts GaitGL-based unsupervised recognition on CASIA-B to 87.4% mean rank-1, surpassing prior unsupervised methods by 8.2% and some fully supervised methods, with the largest gains in the clothing-change (CL) condition, 74.2%. On OUMVLP, the method reaches 68.5% mean rank-1 when GREW is the source dataset.","pith_inferences":["The reported results likely rely on knowing the number of identities in the target training split, 74 on CASIA-B and 5,153 on OUMVLP; on a genuinely unlabeled deployment this number must be estimated, so the practical gain may be smaller than the claimed 8.2 points.","The same dynamic-threshold and confidence-refinement recipe should transfer to person re-identification and other fine-grained biometrics, where clustering-based pseudo-label noise is also the shared bottleneck.","The large source-dataset gap on OUMVLP, 49.2% with a CASIA-B source versus 68.5% with a GREW source, suggests that source diversity, not only label refinement, sets the ceiling; a natural extension is self-supervised pretraining on unlabeled in-the-wild gait video.","A testable prediction is that replacing the teacher branch's body-only augmentation with augmentation that also perturbs head and leg regions would lower CL-condition accuracy, because those regions are nearly unaffected by real clothing changes."],"forward_implications":["On CASIA-B with a GaitGL backbone, GaitDCCR reaches 87.4% mean rank-1, outperforming UGRSF, STANet, and GOUDA, and beating some fully supervised methods, especially in the CL clothing condition.","On OUMVLP as target, the method reports 68.5% mean rank-1 with a GaitSet backbone and 60.6% with GaitGL when GREW is the source, a jump of over 30 points over GOUDA.","Ablation results attribute separate gains to each component: dynamic clustering parameters and dynamic weighted centroids improve clustering quality, while confidence-based pseudo-label refinement and the contrastive teacher module contribute the larger share of the accuracy increase.","The exponential decay schedule for the clustering threshold beats fixed thresholds and other decay curves, indicating that adapting the clustering acceptance criterion during training is a direct source of improvement.","The method works with both GaitSet and GaitGL backbones, suggesting the refinement and dynamic clustering strategies are backbone-agnostic rather than tied to one architecture."],"supporting_citations":[{"why":"Supplies the cluster-based contrastive learning baseline on which the method builds.","marker":"[43]"},{"why":"GaitSet is one of the two backbone feature extractors used in experiments.","marker":"[10]"},{"why":"GaitGL is the backbone used for the main 87.4% CASIA-B result.","marker":"[11]"},{"why":"CASIA-B is the primary benchmark dataset and evaluation protocol.","marker":"[1]"},{"why":"OUMVLP is the large target dataset for the CASIA-B-to-OUMVLP and GREW-to-OUMVLP experiments.","marker":"[46]"},{"why":"STANet is the prior state-of-the-art unsupervised method compared against on CASIA-B.","marker":"[25]"},{"why":"UGRSF is a baseline that also uses memory-bank and clothing-augmentation ideas, which this work refines.","marker":"[24]"},{"why":"GOUDA is the baseline compared against on OUMVLP.","marker":"[23]"},{"why":"Supplies prior data-augmentation ideas for gait silhouettes that motivate the body-region dilation and erosion augmentation.","marker":"[45]"},{"why":"Supplies the teacher-student mutual learning structure used in the contrastive teacher module.","marker":"[26]"}],"fun_headline_variants":["Unsupervised gait recognition hits 87.4% rank-1, +8.2 over prior SOTA","GaitDCCR: unsupervised gait net surpasses supervised on clothing change","No labels needed: gait ID accuracy jumps to 87.4% rank-1","Dynamic clustering refines pseudo-labels, boosting gait recognition to 87.4%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the number of clusters C is fixed correctly before training; the paper never states how C is chosen, and since the target dataset is unlabeled, a deployment would have to guess the number of identities, so the reported 87.4% may depend on knowing that number in advance.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised gait recognition hits 87.4% rank-1, +8.2 over prior SOTA","GaitDCCR: unsupervised gait net surpasses supervised on clothing change","No labels needed: gait ID accuracy jumps to 87.4% rank-1","Dynamic clustering refines pseudo-labels, boosting gait recognition to 87.4%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2872,"prompt_tokens":984,"completion_tokens":1888,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":1793}},"tokens_in":600,"tokens_out":1888,"duration_ms":24616,"temperature":1.0,"reasoning_tokens":1793,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T11:56:34.421942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GaitDCCR on a target set whose true identity count is withheld, estimating C from the feature distribution by silhouette analysis or a density-based cluster count, and report rank-1; if accuracy drops sharply relative to the run that sets C to the known training-split identity count, the central claim of fully label-free adaptation is falsified.","supporting_citations":[{"cited_title":"Takemura, Y","cited_arxiv_id":null,"evidence_quote":"OUMVLP is the large target dataset for the CASIA-B-to-OUMVLP and GREW-to-OUMVLP experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"STANet is the prior state-of-the-art unsupervised method compared against on CASIA-B."},{"cited_title":"Unsupervised Gait Recognition with Selective Fusion","cited_arxiv_id":"2303.10772","evidence_quote":"UGRSF is a baseline that also uses memory-bank and clothing-augmentation ideas, which this work refines."},{"cited_title":"Watch Where You Head: A View-biased Domain Gap in Gait Recognition and Unsupervised Adaptation","cited_arxiv_id":"2307.06751","evidence_quote":"GOUDA is the baseline compared against on OUMVLP."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GaitSet is one of the two backbone feature extractors used in experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CASIA-B is the primary benchmark dataset and evaluation protocol."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the cluster-based contrastive learning baseline on which the method builds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies prior data-augmentation ideas for gait silhouettes that motivate the body-region dilation and erosion augmentation."},{"cited_title":"Zhang, T","cited_arxiv_id":null,"evidence_quote":"Supplies the teacher-student mutual learning structure used in the contrastive teacher module."}],"review_version":1}