{"id":"6a483d74-2c5b-4ef6-a56e-71eee65afa6b","arxiv_id":"2508.03397","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Combining silhouettes with monocular depth estimates from the same RGB frames, fused through a multi-scale cross-level attention network, raises gait recognition rank-1 accuracy on CCPG, SUSTech1K, and CASIA-B.","lead":"This paper adds estimated depth maps, generated from ordinary RGB video by a pretrained model, to the usual silhouette input for gait recognition, and reports accuracy gains on three benchmarks. The result matters because adding a cheap pseudo-depth channel could improve surveillance and person identification without new sensors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth-specific benefit is not isolated: the paper lacks an RGB-only or silhouette-derived control, leaving open that the depth branch merely re-encodes silhouette shape or adds capacity.","rationale":"The reader's weakest_assumption correctly identifies that the depth modality's contribution is not shown to be geometric or complementary beyond silhouette-derived shape. My stress-test converges on the same point: the paper's methodology and ablations never rule out that any second informative input, or even a re-encoding of the silhouette, would produce the same gains. The SUSTech1K night condition and the per-frame normalization make this concern concrete rather than hypothetical. I agree with the conditional verdict: the method may work, but the central mechanistic claim and the SOTA claim need additional control experiments. I would keep the verdict as conditional, not reject, because the reported numbers are internally plausible and the missing control is straightforward to run.","tokens_in":14834,"tokens_out":3697,"duration_ms":51878,"concrete_test":"On CCPG, replace the depth-branch input in DepthGait with (a) the RGB frame masked, aligned, and resized to 64x44 (grayscale), (b) the silhouette's distance transform under the same cropping and normalization, and (c) a spatial permutation of the depth map within the silhouette, keeping the same DeepGaitV2 encoder, MCF fusion, and training settings. If (a) or (b) matches or approaches the 87.6% mean rank-1 accuracy, the depth-specific 3D claim is unsupported; if the real depth map clearly outperforms all three controls, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RGB-derived depth maps provide 3D geometric cues not available in silhouettes and that this explains the state-of-the-art gains. The load-bearing assumption enters in Section 3.2: Depth Anything estimates are converted to disparity, normalized per-frame, masked by the silhouette, and resized to 64x44. After this pipeline, the depth map is highly correlated with the silhouette boundary and interior shape, and per-frame min-max normalization removes absolute metric scale. The experiments never compare against a control branch fed the same RGB pixels (e.g., cropped/aligned grayscale RGB) or a silhouette-derived geometric feature (e.g., a distance transform), so the observed gains in Tables 2 and 5 could come from extra network capacity, a second pathway, or low-level image cues rather than from depth-specific information. The paper's own SUSTech1K night result supports this worry: when silhouette quality is poor, depth accuracy drops to 38.4%, below two peer methods, indicating the depth modality is strongly coupled to silhouette shape rather than acting as an independent geometric signal. The CASIA-B result in Table 4 is depth-only and omits DeepGaitV2, further weakening the claim that the full method achieves state-of-the-art due to depth geometry.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DepthGait, a dual-branch gait recognition network that combines binary silhouette sequences with depth maps estimated from RGB frames using the Depth Anything foundation model. The depth maps are converted to disparity, normalized per frame, masked by the silhouette, and aligned to the silhouette grid; both branches use DeepGaitV2-style encoders, and a multi-scale cross-level fusion (MCF) module with multi-scale attention is applied at every encoder stage. The method is evaluated on CCPG, SUSTech1K, and CASIA-B*, reporting mean rank-1 accuracies of 87.6%, 87.6%, and 91.2% respectively, and the CCPG ablations show monotonic accuracy gains when the depth branch, MSF, and CLF are added.","tokens_in":15024,"tokens_out":7986,"duration_ms":88651,"significance":"If the depth-induced gains are genuine, the work offers a practical way to inject 3D body-geometry cues into silhouette-based gait recognition using existing monocular RGB infrastructure, without LiDAR or multi-view capture. The paper is clearly structured, compares against a broad set of gait modalities and fusion baselines, and its CCPG ablation is clean enough to show that each proposed component helps. The depth-only variant (DepthGait★) also outperforms several established silhouette and skeleton methods, which is encouraging. The main unverified premise is that RGB-derived depth maps contribute geometric information beyond what is already contained in the silhouette and the RGB pixels; the current experiments do not isolate that factor.","major_comments":[{"comment":"The central claim that RGB-derived depth maps provide 3D cues not available in silhouettes is not supported by the current experiments. The depth maps are generated from the same RGB frames, then masked with the silhouette, per-frame min-max normalized in disparity space, and resized to the silhouette grid; after this pipeline the depth input is highly correlated with the silhouette's shape and interior intensity profile. The paper never compares against a control branch that receives the same RGB pixels (e.g., cropped/aligned grayscale RGB), a silhouette distance transform, or a second copy of the silhouette with the same network capacity and fusion. Therefore the improvements in Tables 2 and 5 (e.g., 83.3 to 87.6 in Table 5) could be due to extra capacity, a second pathway, or low-level intensity information rather than depth-specific geometry. The SUSTech1K night result in Table 3 (38.4%, below ParsingGait and SkeletonGait++) is consistent with the depth maps inheriting silhouette failures, as the authors acknowledge. I request a control experiment that removes the depth-specific signal while preserving pathway and capacity.","section":"§3.2, §4.4 (Tables 2 and 5)"},{"comment":"The CASIA-B evaluation is presented as supporting state-of-the-art performance, but Table 4 reports only the depth-only variant ('Depth Map / DepthGait ours') and does not include the full fused DepthGait or DeepGaitV2 in the comparison. The abstract and conclusion refer to the overall method, so Table 4 cannot substantiate that claim. Please report the fused model on CASIA-B and include the missing recent baselines, or explicitly scope the claim to the depth-only variant.","section":"§4.3, Table 4"}],"minor_comments":[{"comment":"The normalization description is internally inconsistent: the text sets the silhouette height to 64 pixels and also says the standardized width w is set to 64, while the implementation later reports an input size of 64×44. Please clarify the exact crop/resize dimensions and whether height or width is fixed.","section":"§3.2"},{"comment":"The final sentence of Section 4.4 promises a visualization analysis comparing depth maps and contour maps, but no such visualization or analysis appears in the paper.","section":"§4.4"},{"comment":"The rows are labeled only with checkmarks under 'Baseline', 'Depth Map', 'MSF', and 'CLF', and the caption does not define whether 'Baseline' denotes the silhouette-only input or the DeepGaitV2 framework. This makes row 2 easy to misread as still containing a silhouette branch; please clarify the notation in the caption.","section":"Table 5"},{"comment":"The incremental gains attributed to MSF and CLF in the text do not match the table: row 2 to row 3 is +2.3/+2.1 (exclude/include), not +3.5/+3.1, and row 3 to row 4 is +0.8/+0.5, not +4.3/+3.6. The cited numbers appear to be cumulative gains over the silhouette baseline rather than incremental gains over the preceding row, so the wording 'further improved' should be corrected.","section":"§4.4, Table 5"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is simple and mostly works: run Depth Anything on the RGB frames, mask with the silhouette, and feed the resulting depth maps alongside the silhouettes into a DeepGaitV2-style network with a multi-scale cross-level fusion. On CCPG the full model gains about four points over silhouette-only, and the ablation is clean: each component (depth input, multi-scale fusion, cross-level fusion) adds a few points. The use of monocular depth foundation models as a gait input modality appears new, and the authors honestly report that performance collapses on SUSTech1K at night.\n\nThe soft spot is the missing control. Depth maps here are a deterministic function of the same RGB pixels that produced the silhouettes, masked to the silhouette region, resized to 64x44, and per-frame min-max normalized. At that scale the depth map is essentially a smooth shading of the silhouette interior. Without a control branch fed the same RGB pixels (e.g., a grayscale crop) or a silhouette-derived geometric feature like a distance transform, the gains could come from extra network capacity, a second pathway, or low-level image cues rather than from depth-specific information. The paper's own night results reinforce this: when silhouettes are poor, depth accuracy drops below several peer methods, suggesting the depth branch is tightly coupled to silhouette quality rather than acting as an independent geometric signal.\n\nThe CASIA-B reporting also overstates the case. Table 4 evaluates a depth-only model, not the full fused DepthGait, yet the abstract and conclusion claim state-of-the-art on that benchmark. DeepGaitV2 is missing from that comparison table. These are addressable issues, but they need fixing.\n\nMinor points: the loss weights and input size are standard free parameters, and the fusion module is a little baroque but ablated. The writing is okay, the references cover the relevant gait literature, and the paper does not hide its failures.\n\nWho should read this: people working on gait recognition or multimodal fusion will find it a useful data point on pseudo-depth as an extra channel. It deserves a serious referee, but the referee should ask for an RGB-only branch or a silhouette-derived geometric control, and for a corrected CASIA-B table that either includes the full model or explicitly labels it as depth-only.","headline":"Novel use of monocular depth for gait recognition, but the paper never isolates whether depth maps add geometric information beyond what an RGB or silhouette-derived control would provide; the CASIA-B state-of-the-art claim rests on a depth-only variant.","tokens_in":15611,"tokens_out":2157,"would_cite":false,"duration_ms":26020,"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":"DepthGait proposes RGB-derived depth maps as a new gait-recognition modality and reports top accuracy on three benchmarks.","keywords":["Gait recognition","Depth maps","Monocular depth estimation","Silhouette sequences","Multimodal fusion","Multi-scale attention","Cross-level fusion"],"falsifier":"Replace the depth map in the pipeline with a silhouette-derived surrogate, such as the distance transform or shaded copy of the silhouette, and retrain under the same fusion scheme; if accuracy matches DepthGait, the depth estimator adds no geometric signal beyond the silhouette, and the paper's central claim fails.","tokens_in":14608,"feed_emoji":"🚶","tokens_out":6655,"duration_ms":76530,"temperature":0.7,"pith_summary":"The paper argues that depth maps estimated from ordinary RGB video carry 3D body-geometry cues, such as stride, limb placement, and torso shape, that binary silhouettes and skeletons miss, and that adding them as a second input modality improves gait recognition under clothing, bag, and viewpoint changes. To test this, the authors build DepthGait, a two-branch network that feeds silhouettes and depth maps through a multi-stage encoder and fuses them at every stage with multi-scale attention and cross-level feature summation. On three benchmarks the fused model reports mean rank-1 accuracy of 87.6% on CCPG, 87.6% on SUSTech1K, and 91.2% on CASIA-B, and a depth-only variant already outperforms silhouette- and skeleton-based peers on CCPG and CASIA-B. The practical point is that existing RGB surveillance footage, without LiDAR or depth sensors, could be retrofitted with a depth-estimation step to strengthen biometric identification.","feed_headline":"Depth maps lift gait recognition past silhouette-only models","feed_subtitle":"A dual-branch network fuses monocular depth with silhouettes, turning 3D body shape into a usable biometric signal.","key_machinery":"The carrying mechanism is the multi-scale cross-level fusion (MCF) module applied at every encoder stage. Given stage features $F_s$ and $F_d$ from the silhouette and depth branches, it concatenates them, runs parallel 1x1-3x3-1x1 and 1x1-5x5-1x1 convolution chains to form local and global scores, and a softmax over the two modality channels produces per-location weights $W_S$; the fused map is $X_f = F_s W_S[:,0] + F_d W_S[:,1]$, and the stage output is $Y_f = X_f + F_{sd}$, where $F_{sd}$ is the encoder's stage output, so fused features feed the next stage. This stage-wise fusion lets low-level geometry from the depth map and appearance from the silhouette reinforce each other progressively, which the ablations show is more effective than single-stage or late fusion.","core_discovery":"The paper's central claim is that depth maps estimated from ordinary RGB frames make gait recognition more accurate than the two dominant input modalities, silhouettes and skeletons, because they encode per-pixel distance to the camera and preserve 3D body geometry that appearance-based masks discard. The preprocessing step converts estimated depth $d$ to normalized disparity $q=1/d$, crops it with the silhouette's bounding box, aligns it on the silhouette's vertical center, and resizes both to 64 by 44 pixels so the two modalities share geometry. DepthGait then runs the two sequences through separate branches of the same multi-stage encoder and, at every stage, fuses the feature maps with multi-scale attention and adds the fused result back into both branches. On the three tested benchmarks the depth-only variant already beats all compared silhouette and skeleton methods on CCPG and CASIA-B, and the fused DepthGait reports the highest mean rank-1 accuracy on all three: 87.6% on CCPG, 87.6% on SUSTech1K, and 91.2% on CASIA-B.","pith_inferences":["Extension not tested in the paper: swapping the monocular depth estimator for a stereo or LiDAR depth source would separate the method's geometric hypothesis from the artifacts of any single estimator; the paper's formulation predicts the gains would persist.","The authors note their own night-condition weakness on SUSTech1K (38.4% rank-1, below two peers), which implies the depth branch inherits silhouette-segmentation failures; a testable fix is jointly refining segmentation and depth rather than masking precomputed silhouettes.","The paper evaluates only the benchmark viewing angles, so a further test of the view-invariance claim would train on a subset of viewpoints and probe unseen angles; if depth geometry is truly view-consistent, the drop should be smaller for depth than for silhouette."],"forward_implications":["Because depth maps are derived from the same RGB frames that already feed silhouette pipelines, existing surveillance datasets can be re-processed with depth estimation and re-evaluated without any new sensor hardware.","The depth-only variant's margins over silhouette and skeleton methods on CCPG and CASIA-B imply that adding a depth branch is a direct accuracy lever wherever RGB video is available.","The ablation results indicate that fusing at every encoding stage, rather than once at a late stage, is what preserves modality-specific detail; late single-stage fusion loses several accuracy points.","The reported person re-identification gains on CCPG (95.2% mean rank-1) suggest the depth-plus-silhouette representation generalizes beyond gait to identity matching from video.","On SUSTech1K the method improves mean rank-1 over the best peer in each category by 6.3%, with the largest gains in normal, bag, and occlusion conditions."],"supporting_citations":[{"why":"Supplies the monocular depth-estimation foundation model that generates the depth maps, the paper's new modality.","marker":"[36]"},{"why":"Provides the multi-stage gait feature encoder reused as both silhouette and depth branches.","marker":"[8]"},{"why":"Defines the CCPG cloth-changing benchmark and its gait and re-identification evaluation protocols used for the main comparisons.","marker":"[19]"},{"why":"Introduces the SUSTech1K benchmark and protocols on which the method reports 87.6% mean rank-1; also the LiDAR-gait baseline the depth-modality argument extends.","marker":"[30]"},{"why":"Supplies the CASIA-B dataset split and evaluation protocol used for the third benchmark.","marker":"[21]"},{"why":"GaitSet is the standard silhouette baseline that the depth-only and fused variants are compared against.","marker":"[4]"},{"why":"SkeletonGait++ is the strongest multimodal peer method and the skeleton-map representation the paper contrasts with depth maps.","marker":"[10]"},{"why":"ParsingGait provides the human-parsing modality baseline and a fusion comparison on the same benchmarks.","marker":"[41]"}],"fun_headline_variants":["Depth maps fused with silhouettes top gait recognition benchmarks","RGB-derived depth gives gait recognition a 3D edge","DepthGait: fusing depth and silhouette cues boosts gait ID","Monocular depth plus silhouettes: a stronger gait signature"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a monocular depth-estimation model, applied to small 64-by-44 crops and masked by the silhouette, recovers genuine 3D body-shape information that is not already contained in the silhouette itself.","fun_headline_variants_meta":{"raw":{"variants":["Depth maps fused with silhouettes top gait recognition benchmarks","RGB-derived depth gives gait recognition a 3D edge","DepthGait: fusing depth and silhouette cues boosts gait ID","Monocular depth plus silhouettes: a stronger gait signature"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001447,"raw_usage":{"total_tokens":5830,"prompt_tokens":948,"completion_tokens":4882,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":4814}},"tokens_in":564,"tokens_out":4882,"duration_ms":38254,"temperature":1.0,"reasoning_tokens":4814,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:29:01.385979+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the depth map in the pipeline with a silhouette-derived surrogate, such as the distance transform or shaded copy of the silhouette, and retrain under the same fusion scheme; if accuracy matches DepthGait, the depth estimator adds no geometric signal beyond the silhouette, and the paper's central claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the monocular depth-estimation foundation model that generates the depth maps, the paper's new modality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the CCPG cloth-changing benchmark and its gait and re-identification evaluation protocols used for the main comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the SUSTech1K benchmark and protocols on which the method reports 87.6% mean rank-1; also the LiDAR-gait baseline the depth-modality argument extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CASIA-B dataset split and evaluation protocol used for the third benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GaitSet is the standard silhouette baseline that the depth-only and fused variants are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SkeletonGait++ is the strongest multimodal peer method and the skeleton-map representation the paper contrasts with depth maps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ParsingGait provides the human-parsing modality baseline and a fusion comparison on the same benchmarks."}],"review_version":1}