{"id":"4eaf4973-7ab9-41e3-a1f5-969129cf063b","arxiv_id":"2412.11495","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new gait recognition model, MultiGait++, fuses silhouette, parsing, and optical flow by separating shared and modality-specific features, and reports state-of-the-art results on four gait benchmarks.","lead":"This paper compares three gait recognition inputs, namely silhouette, body parsing, and optical flow, and proposes a fusion strategy called C2Fusion that separates shared and unique features, building a model named MultiGait++. On four public gait datasets, the model reports the highest rank-1 accuracies, which matters for surveillance and person identification from walking video.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MultiGait++'s reported gain over its MultiGaits+p+f baseline is not capacity-controlled: the C2 module adds a third learnable branch, so observed accuracy gains may reflect extra parameters rather than the C2Fusion mechanism.","rationale":"The reader's weakest_assumption concerns external extractor quality, which affects practical deployment and cross-dataset generalization but does not directly threaten the internal validity of the reported empirical claim: the paper's results are measured under the stated preprocessing. A more load-bearing issue is the capacity confound in the C2Fusion ablation. MultiGait++'s improvement over MultiGaits+p+f is the primary evidence for the proposed fusion strategy, yet the comparison introduces a third learnable branch without controlling for parameters. This is a classic confound that could explain the observed gains and undermines the paper's central technical contribution. The paper even provides a parameter table (Table 7) but only for the full pipeline, not for the ablation variants, so the reader cannot assess whether the gains are due to extra parameters. This issue also affects all four datasets, unlike the SUSTech1K design-selection issue which is limited to one benchmark. Fixing it requires a straightforward control experiment. The concern is not evidence of fraud or false results; it is a missing experimental control, so the appropriate verdict remains CONDITIONAL pending the control. My assessment therefore diverges from the reader's identified weakest_assumption but supports a similar overall verdict, hence 'disagree' on the specific load-bearing concern.","tokens_in":15402,"tokens_out":9028,"duration_ms":80764,"concrete_test":"Train a capacity-matched baseline on SUSTech1K (and verify on CCPG) with the same total parameters and FLOPs as MultiGait++, using the MultiGaits+p+f two-branch architecture but widening channels to reach equal capacity, or alternatively adding a third branch that receives a simple concatenation of f_ap and f_mo (instead of the C2-gated average) with the same Stage2/3 width. If this baseline closes the rank-1 gap to within ~0.5%, the C2Fusion advantage is not supported; if the gap persists, the mechanism is validated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that C2Fusion improves gait recognition depends on the comparison between MultiGait++ (Table 6d) and its baseline MultiGaits+p+f (Table 6a). These differ not only in the C2 fusion logic but also in architecture: MultiGait++ adds a third learnable 'common branch' (f'_co in Eq. 3, Figure 3b) containing Stage2 and Stage3, whereas MultiGaits+p+f fuses only the appearance and motion branches (Figure 3c). This increases parameter count and FLOPs for MultiGait++; the paper does not report a capacity-matched baseline. Consequently, the consistent gains (SUSTech1K R-1 86.0->87.4; CCPG 85.5->87.6; GREW 91.4->93.4; Gait3D 83.0->85.4) could stem from additional model capacity, not from the common/different-concern masking itself. Table 6's ablation removes mco or mdi inside MultiGait++ but never controls for the extra branch; (b) still uses the refined branches without the common branch, while (d) has all components. Without a control that matches parameters/FLOPs, the specific contribution of C2Fusion is not established, weakening the paper's main technical novelty and the causal interpretation of the SoTA results.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies three image-based gait modalities—binary silhouette, human parsing, and optical flow—under a uniform DeepGaitV2 backbone. It introduces the MultiGait series of uni- and multi-modal baselines, evaluates fusion at input, middle, and high levels with concatenation, addition, and attention mechanisms, and draws conclusions about the homogeneity of silhouette+parsing versus the complementarity of silhouette+flow. Based on these observations, the authors propose C2Fusion, a module that extracts a common attention mask across branches and uses its complement to emphasize branch-specific features, leading to a three-branch model called MultiGait++. Experiments on SUSTech1K, CCPG, GREW, and Gait3D report state-of-the-art results, with additional ablations and a parameter/FLOP analysis in the supplementary material.","tokens_in":15818,"tokens_out":5088,"duration_ms":46537,"significance":"If the central claim holds, the paper makes a useful empirical contribution: it provides one of the first fair, apples-to-apples comparisons of three popular gait modalities and shows consistent gains from multimodal fusion on multiple benchmarks. The strength of the paper is its breadth: Table 1 alone contains 18 fusion configurations, the implementation follows official protocols, the code is released, and the supplementary reports a full pipeline cost table (Table 7) rather than only recognition FLOPs. However, the specific novelty of C2Fusion is not yet established: the main comparison is not capacity-controlled, Equation (1) is described as cross-attention despite being per-branch SE gating, and several architecture choices are selected using the same SUSTech1K test set on which the final method is evaluated. These issues are fixable and do not invalidate the empirical study, but they must be addressed before the causal claim about C2Fusion can be accepted.","major_comments":[{"comment":"The headline comparison MultiGait++ (row d) versus its baseline MultiGaits+p+f (row a) is not capacity-matched. MultiGait++ introduces a third learnable common branch that contains Stage2 and Stage3 (Figure 3b), whereas MultiGaits+p+f fuses only the appearance and motion branches (Figure 3c). The ablations in rows (b) and (c) remove the mco or mdi masks but keep the extra branch, so they cannot separate the effect of the C2 masking from the effect of the added parameters and FLOPs. Because the reported gains over the baseline are modest (86.0 to 87.4 on SUSTech1K, 85.5 to 87.6 on CCPG, 91.4 to 93.4 on GREW, 83.0 to 85.4 on Gait3D), a capacity-matched control is required to support the causal claim that C2Fusion, rather than additional model capacity, is responsible for the improvements. A control could be a common branch whose features are combined by plain addition or concatenation without mco/mdi, or a two-branch model matched in parameters and FLOPs.","section":"MultiGait++ / Table 6"},{"comment":"The operation described as 'cross-attention' is in fact two independent squeeze-and-excitation mappings E_ap and E_mo followed by an element-wise softmax; there is no query-key interaction between the two branches. Calling this cross-attention overstates the mechanism and makes the common/different-concerns interpretation less direct. The text and figure should either be changed to 'attention' or 'gating', or a genuine cross-branch attention (for example, one branch's attention map conditioned on the other branch's features) should be used and justified.","section":"Equation (1) and Figure 3(a)"},{"comment":"The fusion-location and fusion-mechanism conclusions (input-level fusion for silhouette+parsing, high-level fusion for silhouette+flow) are read from the SUSTech1K results in Table 1, and the final MultiGait++ model is then evaluated on the same SUSTech1K test protocol in Table 3. This creates a selection-on-test risk for the architecture choices. To support the claim that the design principles generalize, the authors should either use a held-out validation split for choosing fusion locations and mechanisms, or validate the chosen configuration on an independent dataset before reporting final SUSTech1K numbers.","section":"Method / Table 1"},{"comment":"All results are single-run numbers without variance estimates. Differences of one to two rank-1 points (e.g., SUSTech1K 86.0 vs 87.4, or the CL subset 50.4 vs 44.2 in Table 3) are small relative to typical training variance in gait recognition, so the consistency of the gains cannot be assessed. The authors should report means and standard deviations over at least three random seeds, or otherwise provide a justification for why repeated-run variance is negligible under the official protocol.","section":"Experimental reporting / Tables 3-5"}],"minor_comments":[{"comment":"The Introduction refers to 'Figure 4' for the three modalities and the three fusion strategies, but Figure 4 appears in the supplementary material; the main-text figure references should be renumbered consistently (the intended references appear to be Figures 1 and 2).","section":"Introduction"},{"comment":"Table 4 contains typos: 'GaitGrapgh2' should be 'GaitGraph2', and the GaitSet row contains '65,1' instead of '65.1'.","section":"Table 4"},{"comment":"The caption of Table 6 does not define all rows: (a) is called MultiGaits+p+f and (d) MultiGait++, but rows (b) and (c) are described only in the text. Please state in the caption which components are present in each row, especially whether the common branch is retained.","section":"Table 6"},{"comment":"The two-branch variants MultiGait++s+f and MultiGait++s+p used on GREW and Gait3D are not specified precisely. The paper only says the model is 'modified'; please provide the exact input assignment, whether the C2 module and common branch are unchanged, and the training hyperparameters for those datasets.","section":"More Results on Other Real-world Datasets"},{"comment":"The common-branch feature is formed as (f_ap + f_mo)/2 * mco; if f_ap and f_mo have different scales after their respective refinements, the fixed averaging should be justified or replaced by a learned combination.","section":"Equation (3)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical study with broad coverage, and the code release is a plus. The main risk is that the central novelty, C2Fusion, is not isolated from the added common branch in the ablations. I would ask the authors to add a capacity-matched control and to clarify the 'cross-attention' terminology before the paper can be accepted. The selection-on-test concern for Table 1 is also important but can be addressed with a validation split or an additional dataset for model selection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The part of this paper worth keeping is the comparative study. Under a single DeepGaitV2 backbone, the authors systematically test three modalities and a range of fusion locations and mechanisms on the same protocols, and they include a parameter-matched two-stream control (MultiGait2s). That is a real service to the gait community, and the consistency of the gains across SUSTech1K, CCPG, GREW, and Gait3D makes the multimodal argument credible. The code being folded into OpenGait also matters.\n\nThe soft spot is the capacity control in the central claim. Table 6 compares MultiGaits+p+f (two branches) against MultiGait++ (three branches), and the extra branch adds learnable parameters and FLOPs. The ablation toggles the mco and mdi masks, but never holds architecture capacity fixed. So the +1.4 to +2.1 rank-1 improvements could come substantially from the extra branch rather than from the common/different-concern masking. This is not fatal -- the C2 design is plausible and the gains are consistent -- but the paper should have run a wider two-branch baseline or a capacity-matched three-branch control. A referee should ask for this.\n\nThe other concerns are smaller. The fusion-location choices were indeed read off SUSTech1K results, and that same benchmark then reports the final numbers; the transfer to CCPG, GREW, and Gait3D mitigates the circularity, but the SUSTech1K headline number is partly tuned. Calling Eq. 1 \"cross-attention\" is wrong; it is SE-style gating with a softmax over the spatial map, not attention across modalities. Missing error bars are standard for this field, so I would not ding them hard.\n\nThe supposed contradiction about parsing/flow on GREW and Gait3D does not hold up. The main text says those datasets lack the modalities, and the supplement explains they extracted them from RGB themselves. Consistent.\n\nBottom line: this is a solid, useful paper for anyone working on multimodal gait recognition, and the comparative study alone justifies sending it out. The main technical claim needs a capacity-matched ablation, but I would engage with it seriously. Let it go to peer review with a request for that control.","headline":"A genuinely useful unified comparison of silhouette, parsing, and flow for gait, with a plausible fusion mechanism whose specific gain is not yet cleanly separated from added parameters.","tokens_in":16225,"tokens_out":1657,"would_cite":true,"duration_ms":16481,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that fusing silhouettes, body parsing and optical flow via a common-versus-different fusion scheme gives the best reported gait identification accuracy on four benchmarks.","keywords":["gait recognition","multimodal fusion","silhouette","human parsing","optical flow","C2Fusion","soft biometrics","person identification"],"falsifier":"Rerun the SUSTech1K experiment with MultiGait++ against its own naive three-branch concatenation baseline under identical training. If the C2 module does not improve overall top-match accuracy beyond 86.0 to 87.4, the paper's central claim about the fusion strategy fails. A second decisive test is to corrupt the parsing and flow inputs, for example by replacing them with low-resolution or misaligned versions, and check whether the advantage over a silhouette-only system collapses, which would confirm the gains are borrowed from the external extractors.","tokens_in":15188,"feed_emoji":"🚶","tokens_out":12548,"duration_ms":97431,"temperature":0.7,"pith_summary":"This paper tries to establish that combining three image-based gait modalities—binary silhouette, human body parsing, and optical flow—outperforms any single one, and that the fusion rule matters more than the backbone. Under a deliberately uniform architecture and training setup, it reports that silhouette and parsing are similar enough to merge at the input, while silhouette and optical flow carry different information and should be combined at a higher level. From those observations it builds MultiGait++, whose C2Fusion strategy keeps a shared-feature branch and re-weights each modality branch to amplify what is unique to it. It reports the best known top-match identification accuracy on Gait3D, GREW, CCPG, and SUSTech1K at a modest increase in computation over a strong silhouette-only system. If the result holds, surveillance and person-retrieval systems can gain accuracy by adding two additional precomputed image streams rather than redesigning the network.","feed_headline":"Fusing silhouette, parsing and flow sets gait-ID records","feed_subtitle":"A fusion scheme that keeps shared cues and highlights unique ones lifts identity-matching accuracy on four benchmarks.","key_machinery":"The load-bearing component is the C2 module inside C2Fusion. After the appearance and motion branches produce feature maps $f_{ap}$ and $f_{mo}$, two small squeeze-excitation networks project each map to an attention map, and an element-wise softmax aligns them. The element-wise minimum of $m_{ap}$ and $m_{mo}$, followed by min-max normalization, defines a common-concern mask $m_{co}$; its complement $m_{di} = 1 - m_{co}$ marks what is different. The common branch starts from $(f_{ap}+f_{mo})/2$ multiplied by $m_{co}$, while each original branch is reweighted by its own attention times $m_{di}$. Concatenating the three resulting feature streams before the final stages forces the model to retain shared structure and amplify modality-specific signal; the paper's ablations show both masks contribute and that their combination gives the best accuracy.","core_discovery":"On the paper's own terms, the central discovery is that silhouette, human parsing, and optical flow are complementary, and that a fusion rule which explicitly separates shared from unique content can exploit that complementarity. The comparative study shows parsing alone does not beat silhouette, yet adding parsing to silhouette gives a large gain that is nearly independent of fusion mechanism or location; flow alone is competitive with silhouette and contributes most when fused at a high level. MultiGait++ embodies the resulting rules: an appearance branch takes silhouette and parsing concatenated at the input, a motion branch takes optical flow, and a C2 module derives a common-attention mask and a different-attention mask, forming a third common branch while sharpening the modality branches. The paper claims this integrated approach outperforms every compared method on four benchmarks, including the strongest prior multimodal baseline.","pith_inferences":["Because the parsing and flow inputs come from external pretrained extractors, the reported gains are only as good as those extractors; fine-tuning them on the target dataset or training the fusion end-to-end is a natural next test.","The same common-versus-different decomposition could serve as a generic multimodal fusion prior beyond gait—for example, face-plus-body or RGB-plus-depth identification—though the paper only demonstrates it for gait.","The paper's own experiment of appending skeleton maps by input concatenation shows saturation; a more promising route, consistent with its fusion logic, would be to feed the skeleton through a second C2-style module."],"forward_implications":["Adding body parsing and optical flow to a strong silhouette network can raise top-match identification accuracy by several points on clothing-changing and real-world gait datasets, with only about 10–15% extra computation in the recognition stage.","Fusion location is not a free choice: silhouette and parsing, being homogeneous, should be merged at the input, while silhouette and flow, being heterogeneous, should be merged at a high level.","On datasets that lack parsing or flow, the same C2 strategy adapts to two-branch inputs (silhouette plus flow on GREW, silhouette plus parsing on Gait3D) and still reports the best known accuracy.","The common-versus-different decomposition is the active mechanism: removing either mask lowers accuracy, and keeping both outperforms naive concatenation of the same three modalities."],"supporting_citations":[{"why":"Supplies the strong silhouette-based architecture and training recipe that all MultiGait baselines and MultiGait++ reuse.","marker":"(Fan et al. 2024a)"},{"why":"Establishes human parsing as a gait modality with fine-grained part-level shape cues, which the paper adds to silhouettes.","marker":"(Zheng et al. 2023)"},{"why":"Provides evidence that optical flow carries rich motion information for gait, motivating its inclusion as a third modality.","marker":"(Castro et al. 2024)"},{"why":"Prior fusion of silhouette and optical flow; serves as the comparison point for the paper's silhouette-plus-flow results.","marker":"(Feng, Yuan, and Fan 2023)"},{"why":"Contributes the cross-attention fusion mechanism that the C2 module adapts, and the skeleton-map modality used in an additional experiment.","marker":"(Fan et al. 2024b)"},{"why":"Introduces the GREW real-world benchmark and its evaluation protocol for the two-branch silhouette-plus-flow variant.","marker":"(Zhu et al. 2021)"},{"why":"Introduces the Gait3D real-world benchmark and the dense 3D representation behind its protocol.","marker":"(Zheng et al. 2022a)"},{"why":"Defines the CCPG cloth-changing benchmark and its four clothing-condition evaluation protocol.","marker":"(Li et al. 2023)"},{"why":"Provides the SUSTech1K dataset and protocol used for the main fusion-location study and ablation experiments.","marker":"(Shen et al. 2023a)"}],"fun_headline_variants":["Silhouette, parsing, flow: complementary gait signals","Gait ID gains from separating shared and unique cues","Fusing gait modalities by splitting common and distinct","MultiGait++: separate then fuse for better gait ID","Shared vs unique gait cues: fusion insights boost ID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on two external pretrained programs, one that labels body parts and one that estimates motion, to produce its extra input images; if those programs give poor results on a new dataset, the fusion gains will shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Silhouette, parsing, flow: complementary gait signals","Gait ID gains from separating shared and unique cues","Fusing gait modalities by splitting common and distinct","MultiGait++: separate then fuse for better gait ID","Shared vs unique gait cues: fusion insights boost ID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1503,"prompt_tokens":972,"completion_tokens":531,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":453}},"tokens_in":588,"tokens_out":531,"duration_ms":4344,"temperature":1.0,"reasoning_tokens":453,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:51:22.935856+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the SUSTech1K experiment with MultiGait++ against its own naive three-branch concatenation baseline under identical training. If the C2 module does not improve overall top-match accuracy beyond 86.0 to 87.4, the paper's central claim about the fusion strategy fails. A second decisive test is to corrupt the parsing and flow inputs, for example by replacing them with low-resolution or misaligned versions, and check whether the advantage over a silhouette-only system collapses, which would confirm the gains are borrowed from the external extractors.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes human parsing as a gait modality with fine-grained part-level shape cues, which the paper adds to silhouettes."},{"cited_title":"M.; Delgado-Esca \\ n o, R.; Hern \\'a ndez-Garc \\' a, R.; Mar \\' n-Jim \\'e nez, M","cited_arxiv_id":null,"evidence_quote":"Provides evidence that optical flow carries rich motion information for gait, motivating its inclusion as a third modality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior fusion of silhouette and optical flow; serves as the comparison point for the paper's silhouette-plus-flow results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the GREW real-world benchmark and its evaluation protocol for the two-branch silhouette-plus-flow variant."}],"review_version":1}