betaFeedback

WorksComputer Science

PolarGait: Constructing Complementary Views from a Single Human Sketch for Gait Recognition

Abstract

Existing multimodal gait recognition methods exploit complementary information by fusing heterogeneous representations. Whether useful complementarity can be constructed from a single representation before fusion remains less explored. We propose PolarGait, which constructs complementary views of Human Sketch through filtering and spatial reparameterization. Filter-Oriented Sketch Decomposition (FOSD) uses Gaussian smoothing to extract coherent structures and absolute residuals to capture local response differences. Gait-Oriented Polar Transformation (GOPT) combines silhouette-guided sketch normalization, a gait-oriented center, and lower-body-biased non-uniform angular sampling to construct a polar view complementary to the Cartesian representation. Silhouette provides auxiliary global body-shape information. Experiments on CCPG, SUSTech1K, CASIA-B, and Gait3D validate the framework. Under the single-Sketch setting, FOSD+GOPT improves the original Sketch baseline by 6.6% on CCPG and 4.6% on CASIA-B. The complete PolarGait achieves Rank-1 accuracies of 87.9%, 95.2%, 96.6%, and 86.4% on the four benchmarks, respectively. These results show that complementary gait information can be constructed within a single structured representation and integrated for robust gait recognition.

Introduction

Gait recognition identifies individuals from their walking patterns and has attracted attention due to its non-contact, long-distance, and non-cooperative capabilities. Unlike face, fingerprint, and iris recognition, gait can be captured at a distance without subject cooperation, making it suitable for surveillance and real-world applications [1].

Vision-based gait recognition has achieved substantial progress on both controlled and in-the-wild benchmarks, partly

driven by the continuous exploration of visual representations derived from RGB videos. Raw RGB images preserve rich appearance and contextual information but are sensitive to clothing textures, illumination variations, and background interference [2, 3]. Silhouette suppresses most appearance information and captures global body shape, but discards internal structural details [4, 5, 6]. Human Parsing further decomposes the body into semantic parts, providing richer internal structures while remaining dependent on upstream parsing quality and label granularity [7, 8]. Skeleton represents the human body through articulated joints and their spatial relationships, providing compact structural and motion cues while relying on the accuracy of pose estimation [9, 10]. These representations encode gait information from different perspectives and exhibit complementary characteristics, motivating their integration in multimodal gait recognition.

As illustrated in Figure 1(a), existing multimodal gait recognition methods mainly exploit complementary information across heterogeneous gait representations. BiFusion [11], MM-GaitFormer [12], SkeletonGait++ [10], and TriGait [13] combine Silhouette and Skeleton to jointly model global body shape and articulated structural information. XGait [14] integrates Silhouette with Human Parsing to exploit their complementary structural properties at different granularities. Multi-Gait++ [15] and TriFusionGait [16] further extend multimodal gait recognition to trimodal settings by incorporating Optical Flow and Skeleton, respectively, thereby introducing additional motion or articulated structural information into the fusion process. SketchGait [17] combines Human Sketch and Human Parsing to exploit their complementary structural characteristics. More recently, GaitFuse [18] further considers cross-modal semantic discrepancy and modality reliability during multimodal fusion. These studies demonstrate that heterogeneous gait representations can provide complementary shape, structural, semantic, and motion information, and that effective fusion depends not only on combining multiple inputs, but also on how information from these representations is effectively interacted, aligned, and integrated. Accordingly, recent multimodal gait methods have increasingly focused on designing more effective interaction and fusion mechanisms for heterogeneous representations. However, the complementary information exploited by existing methods is still primarily provided by heterogeneous input representations, such as the body-shape information from Silhouette and the articulated structural information from Skeleton. It remains less explored whether a single gait representation can be reorganized into multiple views that emphasize different information and subsequently provide complementary cues for gait recognition. This motivates a further question: Can useful complementarity also be actively constructed within a single representation?

Figure 1. Comparison of complementary gait representation learning. (a) Complementarity across heterogeneous gait representations. (b) Complementarity constructed within Human Sketch through filtering and spatial reparameterization.

Human Sketch provides a suitable basis for constructing complementary gait representations. As a dense edge-based representation, it preserves both external body contours and internal structural responses without explicit semantic part labels [17]. Compared with Silhouette, Human Parsing, and Skeleton, Human Sketch contains a wider range of structural information, from coherent body contours to fine-grained local details. However, existing Sketch-based methods [17] primarily exploit complementarity by combining Sketch with other representations, leaving the potential for constructing complementary views within Sketch itself underexplored.

As illustrated in Figure 1(b), we explore two ways to construct complementary views from a single Sketch: filter-based decomposition and spatial reparameterization. First, filter-based decomposition exploits how spatial filtering affects the information within a Sketch. Spatial filtering, such as Gaussian smoothing, preserves recognizable body contours and major structural lines while suppressing fine-grained edge responses. This allows us to construct two complementary views: a smoothed representation retaining coherent body structures and an absolute residual capturing response changes induced by filtering.

Second, spatial reparameterization provides another source of complementarity. Inspired by the Spatial Polar Transformation (SPT) [19] for person re-identification, we explore whether Sketch can be reorganized under a different spatial parameterization. Cartesian and polar spaces describe the same structures through different spatial relationships: the former preserves horizontal and vertical locality, whereas the latter organizes them by radial and angular relations around the body. This reorganization changes how the same Sketch responses are spatially arranged, allowing Cartesian and polar views to capture complementary structural patterns without introducing additional visual information.

Building on this analysis, we propose PolarGait to construct complementary views from Human Sketch. First, Filter-Oriented Sketch Decomposition (FOSD) applies Gaussian smoothing and absolute residual extraction to produce two response views: a smoothed representation preserving coherent structures and a residual representation capturing local response variations. Second, Gait-Oriented Polar Transformation (GOPT) reorganizes Sketch structures in gait-oriented polar space. To accommodate frame-wise variations and gait-specific body characteristics, GOPT employs silhouette-guided sketch normalization (SGSN) to reduce location and scale variations, a gait-oriented center (GOC) for polar unfolding, and lower-body-biased non-uniform angular sampling (LBB-NAS) to emphasize lower-limb gait dynamics. PolarGait then employs three independent branches. The structure branch combines the smoothed Sketch with Silhouette to incorporate global body-shape cues, while the residual and polar branches process their respective representations. Their high-level features are fused for gait recognition.

Extensive experiments are conducted on CCPG, SUSTech1K, CASIA-B, and Gait3D. Under the single-Sketch setting, FOSD+GOPT improves the original Sketch baseline by 6.6% on CCPG and 4.6% on CASIA-B, demonstrating the benefit of constructing complementary representations within Human Sketch. With additional global body-shape information from Silhouette, PolarGait achieves Rank-1 performance of 87.9%, 95.2%, 96.6%, and 86.4% on CCPG, SUSTech1K, CASIA-B, and Gait3D, respectively.

The main contributions of this work are summarized as follows:

  • We demonstrate that complementary gait information can be constructed within a single Human Sketch, rather than relying solely on heterogeneous representations.

  • We propose PolarGait, which constructs complementary Sketch views through FOSD and GOPT. FOSD reorganizes responses via Gaussian smoothing and residual extraction, while GOPT performs gait-oriented polar reparameterization. Silhouette additionally provides global body-shape cues.

  • Experiments on CASIA-B, SUSTech1K, CCPG, and Gait3D demonstrate the effectiveness of the constructed views and strong performance across diverse benchmarks.

Visual Representations for Gait Recognition

Vision-based gait recognition has explored diverse visual representations to capture identity-related information. Silhouette describes global body shape while suppressing appearance details. GaitSet [4] employs set-level aggregation, GaitPart [5] models part-level motion patterns, and GaitGL [6] combines global and local feature learning. TrackletGait [20] further handles unconstrained sequences through tracklet sampling and wavelet-based downsampling. However, Silhouette discards internal structures and fine-grained details. Raw RGB retains richer appearance and contextual information but is sensitive to clothing, illumination, and background variations. DenoisingGait [2] suppresses appearance interference, while BigGait [3] learns gait representations directly from RGB. BiggerGait [48] leverages large-scale pretrained features, and Gait-X [21] improves gait-related feature extraction from appearance-rich inputs. Human Parsing provides explicit body-region semantics. ParsingGait [7] and GaitParsing [8] exploit semantic parts for fine-grained structural modeling, while MultiGait++ [15] and TriFusionGait [16] integrate Parsing with other representations to exploit complementary information. Skeleton represents articulated joints and their spatiotemporal relationships. GaitGraph [9] and GaitGraph2 [22] model joint topology and motion using graph convolutions. Gait-TR [49] combines spatial transformation with temporal convolution, while GaitPT [23] employs hierarchical Transformers. SkeletonGait [10] converts joints into skeleton maps. Despite its reduced sensitivity to appearance, Skeleton depends on pose estimation accuracy and retains limited body-shape information.

More recently, Human Sketch has emerged as a dense edge-based representation preserving body contours and internal structural responses without semantic labels [17]. Unlike other representations, Sketch captures both coherent body structures and fine local details within a unified representation. However, existing Sketch-based methods mainly exploit complementarity by combining Sketch with additional modalities such as Human Parsing, leaving complementary view construction within Sketch itself underexplored.

Spatial Reparameterization for Human Representation Learning

Spatial reparameterization reorganizes visual structures beyond conventional Cartesian coordinates. Spatial Transformer Networks (STN) [24] introduce differentiable spatial transformations, while Active Convolution [50] and Deformable Convolutional Networks [25] learn adaptive sampling locations. Polar Transformer Networks (PTN) and log-polar representations [26] further organize visual information in radial and angular coordinates to model rotation and scale variations.

For human representation, Spatial Polar Transformation (SPT) [19] transforms human contour sketches into polar space for discriminative curve learning in person re-identification. However, generic polar transformation does not account for frame-wise variations in body location and scale or the uneven distribution of gait-related motion cues across body regions. These limitations motivate gait-oriented spatial reparameterization that provides a complementary structural view under gait-adapted polar geometry.

Complementary Representation Learning for Gait Recognition

Information fusion integrates evidence from multiple sources while addressing uncertainty and heterogeneity [27]. In gait recognition, complementary learning primarily combines heterogeneous representations. BiFusion [11] and SkeletonGait++ [10] exploit complementary body-shape and articulated cues from Silhouette and Skeleton, while MM-GaitFormer [12] performs spatial and temporal fusion. GaitMA [29] introduces cross-modal alignment and interaction, whereas TriGait [13] adopts a multi-branch architecture. DAGait [30] further uses Skeleton to guide Silhouette alignment.

Recent methods explore richer modality combinations. XGait [14] integrates Silhouette and Human Parsing through cross-granularity interaction, while SketchGait [17] combines Human Sketch and Human Parsing to exploit complementary structural cues. MultiGait++ [15] combines Silhouette, Human Parsing, and Optical Flow to model shared and modality-specific information, while TriFusionGait [16] incorporates Skeleton for additional structural cues. GaitFuse [18] addresses semantic discrepancy and modality reliability through hierarchical alignment and uncertainty-aware fusion.

These methods mainly obtain complementary information from heterogeneous input representations. In contrast, we construct complementary response and spatial views within a single Human Sketch representation, enabling complementary representation learning before multimodal fusion.

Method

Overview of PolarGait

As illustrated in Figure 2(a), PolarGait constructs complementary gait representations centered on Human Sketch through response reorganization and spatial reparameterization. FOSD produces a smoothed representation with more coherent structural responses and an absolute residual representation that captures local response changes induced by smoothing. SGPT reorganizes the same Sketch in a gait-oriented polar space, providing a complementary structural description under a different spatial reparameterization.

Figure 2. Overview of PolarGait. (a) Three-branch framework with DeepGaitV2 [31] backbones, following [4] for MTP, HPP, and FCs, with BNNeck [33]. (b) Filter-Oriented Sketch Decomposition (FOSD). (c) Gait-Oriented Polar Transformation (GOPT), where LBB-NAS performs angular unfolding in the order π→π/2→0→−π/2→−π\pi\rightarrow\pi/2 \rightarrow0 \rightarrow-\pi/2 \rightarrow-\pi.

Based on these representations, PolarGait employs three independent branches. The structure branch combines the smoothed Sketch with the corresponding Silhouette to incorporate complementary global body shape information. The residual branch models local response variations induced by smoothing, while the polar branch learns structural information from the gait-oriented polar representation.

Filter-Oriented Sketch Decomposition

a) Motivation: Human Sketch contains coherent body structures alongside dense local responses from clothing boundaries, wrinkles, and textures. These responses behave differently under spatial filtering: coherent structures remain recognizable after smoothing, whereas fine-grained patterns undergo greater local changes. This distinction motivates constructing complementary views through response reorganization.

b) Implementation: As shown in Figure 2(b), we propose Filter-Oriented Sketch Decomposition (FOSD) to construct smoothed and absolute residual views from the same Sketch. Given a Sketch sequence X={Xt}t=1TX = \{X_t\}_{t=1}^{T}, the smoothed representation is defined as

St=Gσ(Xt),(1)S_t = \mathcal{G}_{\sigma}(X_t), \tag*{(1)}

where Gσ(⋅)\mathcal{G}_{\sigma}(\cdot) denotes Gaussian convolution with standard deviation σ\sigma. We use a 5×55 \times5 Gaussian kernel for Sketch inputs resized to 128×128128 \times128 pixels.

The absolute residual representation is defined as

Rt=∣Xt−St∣.(2)R_t = \lvert X_t - S_t \rvert. \tag*{(2)}

Unlike direct subtraction, the absolute difference captures both positive and negative response changes induced by smoothing. The smoothed view emphasizes coherent structural responses, while the residual view captures local response variations. Their independent encoding enables complementary information to be exploited for gait recognition.

Gait-Oriented Polar Transformation

a) Motivation: Beyond response reorganization, we construct complementary Sketch views through spatial reparameterization. Unlike Cartesian coordinates that preserve horizontal and vertical locality, polar coordinates organize structures through radial and angular relations [19, 26]. A fixed convolution window in polar space also corresponds to a radius-dependent region in Cartesian space, with tangential extent increasing with radius. This spatial reorganization enables the same Sketch to be represented through different structural relationships and spatial scales.

However, generic polar transformation does not account for frame-wise variations in body location and scale or the uneven distribution of gait-related cues. In particular, the upper body often contains denser appearance-related responses, while the lower limbs exhibit more pronounced gait dynamics. We therefore propose Gait-Oriented Polar Transformation (GOPT), which incorporates silhouette-guided normalization, a gait-oriented center, and lower-body-biased non-uniform angular sampling to adapt polar reparameterization to gait sequences.

b) Implementation: Silhouette-Guided Sketch Normalization (SGSN): As shown in Figure 2(c), given a Sketch frame XtX_t and its corresponding silhouette mask MtM_t, we estimate the vertical foreground range and horizontal mass center from MtM_t. A square cropping region centered on the body is then applied identically to both XtX_t and MtM_t. The cropped Sketch and silhouette are resized to 128×128128 \times128 pixels, producing the normalized inputs X~t\widetilde{X}_t and M~t\widetilde{M}_t, respectively. This normalization reduces variations in body location and scale across frames while preserving thin Sketch structures for subsequent polar resampling.

Gait-Oriented Center (GOC): The polar center determines how body structures are unfolded and arranged in polar space. Although SGSN aligns the body spatially, the fixed center of the 128×128128 \times128 image may not provide an optimal origin for gait-oriented polar unfolding. We define the GOC according to the normalized silhouette:

ctx=Mx(M~t),cty=y~tmin⁡+α(y~tmax⁡−y~tmin⁡),(3)c_t^x = \mathcal{M}_x(\widetilde{M}_t), \qquad c_t^y = \widetilde{y}_t^{\min} + \alpha\left(\widetilde{y}_t^{\max} - \widetilde{y}_t^{\min}\right), \tag*{(3)}

where Mx(⋅)\mathcal{M}_x(\cdot) denotes the horizontal mass center, [y~tmin⁡,y~tmax⁡][\widetilde{y}_t^{\min},\widetilde{y}_t^{\max}] denotes the vertical foreground range of M~t\widetilde{M}_t, and α\alpha controls the normalized vertical position of the center.

Lower-Body-Biased Non-Uniform Angular Sampling (LBB-NAS): Given ct=(ctx,cty)c_t = (c_t^x, c_t^y), each polar coordinate (r,θ)(r,\theta) is mapped to a sampling location in X~t\widetilde{X}_t by

x=ctx+rcos⁡θ,y=cty+rsin⁡θ.(4)x = c_t^x + r\cos\theta,\qquad y = c_t^y + r\sin\theta. \tag*{(4)}

Before angular sampling, we first determine the radial sampling range. Given the normalized 128×128128 \times128 Sketch, the initial sampling radius is set to r=64r = 64 pixels. Since the gait-oriented center is vertically shifted by α\alpha and may deviate from the image center, we extend the sampling radius by a padding margin of 15 pixels, yielding R=r+15=79R = r + 15 = 79 pixels, to ensure sufficient coverage of the entire body.

Uniform angular sampling assigns equal sampling density to different body regions, despite their unequal contributions to gait dynamics. To better capture lower-limb motion, LBB-

InputMethodCCPG CLCCPG UPCCPG DNCCPG BGCCPG MeanSUSTech1K NMSUSTech1K BGSUSTech1K CLSUSTech1K CRSUSTech1K UMSUSTech1K UNSUSTech1K OCSUSTech1K NTSUSTech1K OA@R-1
DepthSimpleView [34]–––––72.368.857.263.349.279.762.566.564.8
LidarGait [35]–––––91.888.674.689.067.594.580.990.486.8
Point CloudHMRNet [36]–––––92.792.379.690.383.195.286.290.490.2
PointNet++ [37]–––––82.578.758.776.174.085.475.874.877.1
LidarGait++ [38]–––––94.293.979.792.491.596.691.992.292.7
SilhouetteGaitSet [4]60.265.265.168.564.869.168.237.465.063.161.067.223.065.0
GaitPart [5]64.367.868.671.768.162.262.833.159.557.254.857.221.759.2
GaitGL [6]–––––67.166.235.963.361.658.166.617.963.1
GaitBase [32]71.675.076.878.675.581.577.549.675.875.576.781.425.976.1
DeepGaitV2 [31]78.684.880.789.283.386.582.849.280.483.381.986.028.080.9
GaitCLIF [39]–––––89.286.658.084.087.788.090.729.984.7
TrackletGait-64 [20]–––––––––––––86.2
SkeletonGaitGraph2 [22]5.05.35.86.25.622.218.26.818.613.419.227.316.418.6
SkeletonGait [10]40.448.553.061.750.967.963.536.561.658.167.279.150.163.0
ParsingDeepGaitV2 [31]68.873.680.187.877.693.187.940.988.887.788.793.547.587.5
Sketch-TDeepGaitV2 [31]61.067.475.783.171.295.292.157.492.390.890.697.961.291.3
FOSD+GOPT77.385.982.693.384.896.995.571.994.894.593.998.869.594.4
Sketch-MDeepGaitV2 [31]70.378.781.389.880.095.593.649.793.392.790.797.567.492.1
FOSD+GOPT79.288.584.794.086.696.495.363.294.694.392.798.270.193.8
silh.+Skel.BiFusion [11]62.667.666.366.065.669.862.345.460.954.363.577.833.762.1
SkeletonGait++ [10]79.183.981.789.983.785.182.946.681.980.882.586.247.581.3
PSGait [40]82.585.388.689.186.487.985.253.984.586.385.391.447.584.6
GaitFuse [18]-----86.383.447.482.681.583.787.447.283.5
silh.+Pars.XGait [14]73.975.579.881.277.6–––––––––
silh.+Pars.+FlowMultiGait++ [15]83.989.086.091.587.692.089.450.487.689.789.193.445.187.4
Sketch+Pars+silh.SketchGait++ (TEED) [17]82.287.086.894.587.697.594.460.193.694.193.998.954.093.1
Sketch+silh.PolarGait (TEED)80.188.583.394.586.697.496.574.495.795.594.898.974.195.2
PolarGait (MUGE)82.190.285.294.287.996.896.264.895.595.394.198.574.894.7

Table 1. Rank-1 accuracy (%) on CCPG and SUSTech1K. CCPG follows the gait recognition protocol, with Mean denoting the average over four conditions. OA@R-1 denotes overall rank-1 accuracy on SUSTech1K. Best results are in bold; “–” denotes unavailable results.

NAS allocates more angular samples to the lower body. Given NθN_{\theta} angular samples, we define

Nlower=round⁡(βNθ),Nupper=Nθ−Nlower,(5)\begin{aligned} N_{\mathrm{lower}} &= \operatorname{round}(\beta N_{\theta}),\\ N_{\mathrm{upper}} &= N_{\theta} - N_{\mathrm{lower}}, \tag*{(5)} \end{aligned}

where β\beta controls the proportion of samples allocated to the lower body. This design provides finer angular resolution for lower-limb structures while retaining the overall upper-body structure with fewer samples. The resulting polar representation is resized to 64×6464 \times64 pixels.

Complementary Multi-View Representation Learning

Based on the complementary views constructed by FOSD and GOPT, PolarGait adopts three independent branches with DeepGaitV2 [31] backbones. Each branch employs an independent gait head comprising Max Temporal Pooling (MTP), Horizontal Pyramid Pooling (HPP), Separate Fully Connected Layers (FCs) [4, 32], and BNNeck [33]. Let S, R, P, and M denote the smoothed Sketch, residual Sketch, polar representation, and normalized Silhouette, respectively. The final gait embedding is obtained through late concatenation:

ePolarGait=Concat⁡(FS([S,M]c),Fr(R),Fp(P)),(6)e_{\mathrm{PolarGait}} = \operatorname{Concat}\left(\mathcal{F}_{S}\left([S,M]_{c}\right),\mathcal{F}_{r}(R),\mathcal{F}_{p}(P)\right), \tag*{(6)}

where [⋅,⋅]c[\cdot,\cdot]_{c} denotes channel-wise concatenation and FS\mathcal{F}_{S}, Fr\mathcal{F}_{r}, and Fp\mathcal{F}_{p} represent the independent branch encoders and gait heads. Classification logits are likewise concatenated for supervision. This late fusion integrates complementary structural and spatial information while preserving branch-specific representations.

Training Objective

PolarGait is optimized using cross-entropy loss and triplet loss [32, 41] to jointly learn discriminative identity representations across the three branches:

L=Lce+Ltri,(7)\mathcal{L} = \mathcal{L}_{\mathrm{ce}} + \mathcal{L}_{\mathrm{tri}}, \tag*{(7)}

where Lce\mathcal{L}_{\mathrm{ce}} supervises identity classification and encourages features from different views to remain discriminative for subject recognition. The triplet loss Ltri\mathcal{L}_{\mathrm{tri}} further constrains the embedding space by reducing intra-class variation and increasing inter-class separation.

Experiments

Experimental Setup

Datasets and Evaluation Protocols

  1. Datasets and Evaluation Protocols: We evaluate PolarGait on four gait recognition benchmarks: CCPG,

InputMethodNMBGCLMeanRank-1Rank-5mAPmINP
SkeletonGaitGraph2 [22]82.073.263.672.9––––
GaitGraph [9]––––6.316.25.22.4
SilhouetteGaitSet [4]96.190.870.385.736.758.330.017.3
GaitPart [5]96.291.578.788.828.247.621.612.4
GaitGL [6]97.494.583.691.829.748.522.313.3
GaitBase [32]97.694.077.489.764.6–––
DeepGaitV2 [31]–––89.674.488.065.839.2
GaitCLIF [39]––––76.588.567.9–
TrackletGait-64 [20]––––77.889.070.2–
ParsingGaitBase [32]––––73.888.666.4–
ParsingGait [7]––––76.289.168.241.3
Sketch (MUGE)DeepGaitV2 [31]99.598.276.591.4––––
FOSD+GOPT99.899.089.296.0––––
Sketch (Parsing)DeepGaitV2 [31]––––82.792.476.159.0
Silh.+Skel.BiFusion [11]98.796.092.095.6––––
SkeletonGait++ [10]––––77.689.470.342.6
GaitFuse [18]––––81.290.871.243.5
Silh.+Pars.XGait [14]––––80.591.973.355.4
Silh.+Pars.+FlowMultiGait++ [15]––––85.494.980.565.2
Sketch+Silh.PolarGait (MUGE)99.999.290.696.6––––
PolarGait (Parsing)––––86.493.780.764.4

Table 2. Rank-1 accuracy (%) on CASIA-B and retrieval performance (%) on Gait3D. Sketch (Parsing) denotes a sketch representation derived from human parsing. PolarGait (Parsing) uses this representation without FOSD.

SUSTech1K, CASIA-B, and Gait3D, covering controlled indoor, challenging outdoor, and in-the-wild scenarios.

CCPG [42] contains 200 subjects under seven clothing conditions and up to ten viewpoints. Following the official protocol, 100 subjects are used for training and 100 for testing. We report Rank-1 accuracy under full clothing change (CL), upper-body change (UP), lower-body change (DN), and backpack (BG), along with their mean rank-1 accuracy.

SUSTech1K [35] contains 1,050 subjects and 25,239 sequences across eight conditions: normal walking (NM), bag carrying (BG), clothing variation (CL), carrying objects (CR), umbrella (UM), uniform (UN), occlusion (OC), and nighttime (NT). Following the official split, 250 subjects are used for training and 800 for testing. We report condition-wise Rank-1 accuracy and overall Rank-1 accuracy (OA@R-1).

CASIA-B [43] contains 124 subjects recorded from 11 viewpoints, each with six normal walking (NM), two bag-carrying (BG), and two clothing-changing (CL) sequences. We use 74 subjects for training and 50 for testing. Following the standard protocol, NM#1–4 form the gallery, while NM#5–6, BG#1–2, and CL#1–2 serve as probes. We report average cross-view Rank-1 accuracy, excluding identical-view matching.

Gait3D [44] contains 4,000 subjects and 25,309 sequences captured by 39 cameras under unconstrained conditions. Following the official protocol, 3,000 subjects (18,940 sequences) are used for training and 1,000 for testing. Evaluation metrics include Rank-1, Rank-5, mean Average Precision (mAP), and mean Inverse Negative Penalty (mINP).

Sketch Generation and Preprocessing

  1. Sketch Generation and Preprocessing: For Sketch generation, we adopt TEED [45] following SketchGait [17] and additionally employ MUGE [46] to evaluate generalization across different edge extractors.

Implementation Details

For SUSTech1K and CASIA-B, the provided silhouettes are not aligned with the original RGB frames. We therefore use SAM3 [47] to generate aligned foreground masks and remove background pixels before Sketch extraction. Due to the low resolution of CASIA-B, we employ Real-ESRGAN [51] to upscale RGB frames by a factor of four before Sketch extraction.

Since Gait3D does not provide original RGB frames, we follow SketchGait [17] to derive coarse Sketches from Human Parsing using morphological dilation and erosion, termed parsing-to-edge.

  1. Implementation Details: All models are trained using SGD with momentum 0.9, weight decay 5×10−45 \times10^{-4}, and an initial learning rate of 0.1, decayed by a factor of 0.1 at predefined milestones. For CASIA-B, CCPG, and Gait3D, training runs for 120K iterations with learning rate decay at 40K, 80K, and 100K, using a batch configuration of 8×8×308 \times8 \times30. For SUSTech1K, training runs for 50K iterations with decay at 20K, 30K, and 40K, using 8×8×108 \times8 \times10. The batch configuration denotes identities ×\times sequences per identity ×\times frames per sequence.

Comparison with State-of-the-Art Methods

Single-Sketch Performance

  1. Single-Sketch Performance: We evaluate FOSD+GOPT using only Sketch as the encoded representation, with Silhouette employed solely for geometric normalization in GOPT. This setting allows us to evaluate the effectiveness of complementary view construction within Human Sketch without additional Silhouette feature fusion.

As shown in Tables I and II, FOSD+GOPT consistently improves the original Sketch baseline under the same DeepGaitV2 [31] backbone. On SUSTech1K, OA@R-1 increases from 91.3% to 94.4% for TEED and from 92.1% to 93.8% for MUGE, with TEED gaining 14.5% and 8.3% under CL and NT, respectively. On CCPG, the mean Rank-1 accuracy improves by 13.6% and 6.6% for TEED and MUGE, respectively. On CASIA-B, MUGE improves from 91.4% to 96.0%.

Compared with the strongest existing single-modal methods, FOSD+GOPT achieves gains of 1.7%, 3.3%, and 4.2% on SUSTech1K, CCPG, and CASIA-B, respectively. These results demonstrate the effectiveness of constructing complementary response and spatial views from a single Sketch representation.

Overall Performance of PolarGait

We evaluate the complete PolarGait by incorporating Silhouette into the smoothed structure branch to complement the constructed Sketch views. As shown in Tables I and II, this integration further improves performance over FOSD+GOPT.

On SUSTech1K, PolarGait achieves OA@R-1 of 95.2% and 94.7% with TEED and MUGE, respectively, improving FOSD+GOPT by 0.8% and 0.9%. The TEED variant also exceeds the previous best result by 2.1%. On CCPG, PolarGait (MUGE) achieves 87.9% mean Rank-1 accuracy, surpassing FOSD+GOPT by 1.3% and MultiGait++ by 0.3%. On CASIA-B, PolarGait reaches 96.6% mean Rank-1 accuracy, a further 0.6% gain.

On Gait3D, where RGB frames are unavailable, we employ parsing-to-edge Sketches and construct a complementary polar view using GOPT. Combined with Silhouette, PolarGait achieves 86.4% Rank-1, 93.7% Rank-5, 80.7% mAP, and 64.4% mINP, attaining the highest Rank-1 accuracy among the compared methods. These results demonstrate the effectiveness of integrating constructed Sketch views with global body-shape information across diverse gait benchmarks.

Analysis of Constructed Complementarity

We investigate the effectiveness of the complementary views constructed by FOSD and GOPT, as well as the contribution of additional Silhouette information. Unless otherwise specified, all ablation experiments are conducted on CCPG using MUGE Sketches with the same backbone and training settings. The corresponding results are reported in Table III.

SketchFOSD SmoothedFOSD ResidualGOPTSilh. FusionCLUPDNBGMeanΔ\Delta Mean
✓–70.378.781.389.880.0–
✓ ×2–73.682.483.292.082.8+2.8
✓–64.474.878.689.076.7-3.3
✓–65.476.178.287.276.7-3.3
✓✓–76.184.184.192.684.2+4.2
✓–67.478.766.682.873.9-6.1
✓✓✓–78.287.481.692.685.0+5.0
✓✓✓–79.288.584.794.086.6+6.6
✓✓✓Residual81.889.984.994.387.7+7.7
✓✓✓Smoothed82.190.285.294.287.9+7.9

Table III. Ablation study of the main components on CCPG in terms of Rank-1 accuracy (%). Sketch ×2 denotes a dual-branch baseline with identical original Sketch inputs. Silh. Fusion indicates the FOSD branch with which the Silhouette is concatenated.

Analysis of FOSD

We first evaluate the effectiveness of FOSD in constructing complementary Sketch views. As shown in Table III, FOSD improves the mean Rank-1 accuracy from 80.0% to 84.2% on CCPG, achieving a gain of 4.2%. In contrast, the smoothed and residual views each achieve only 76.7% when used independently. Their superior combined performance suggests that the two views capture complementary discriminative information rather than simply providing stronger individual representations.

To distinguish the benefit of view construction from increased model capacity, we further introduce a Sketch ×2 baseline, where two independent branches process identical original Sketch inputs. Under the same network configuration and training settings, this baseline achieves 82.8%, improving the original Sketch by 2.8%. FOSD further increases the accuracy to 84.2%, outperforming Sketch ×2 by 1.4%. This result indicates that the gain of FOSD cannot be attributed solely to increased model capacity, providing further evidence for the effectiveness of constructing complementary response views.

Analysis of GOPT

We analyze the gait-oriented center (GOC) and lower-body-biased non-uniform angular sampling (LBB-NAS) on CCPG, with other settings fixed. For GOC, we first compare four polar center strategies: image center, bounding-box center, Silhouette centroid, and the proposed gait-oriented center. Unlike the image-centered SPT [19], GOC determines the polar origin based on body geometry to better accommodate gait-specific structural characteristics. As shown in Table IV, GOC achieves the highest mean Rank-1 accuracy of 69.3%, exceeding the image, bounding-box, and centroid alternatives by 2.0%, 1.6%, and 0.6%, respectively. The gains are most evident under CL and UP.

CenterCLUPDNBGMean
Image center60.765.168.674.667.3
BBox center60.766.768.075.667.7
Silhouette centroid61.667.069.477.068.7
GOC (α=0.60\alpha= 0.60)62.570.167.776.969.3

Table IV. Comparison of polar center strategies on CCPG under uniform angular sampling. Rank-1 accuracy (%).

We further vary α\alpha from 0.58 to 0.62 under uniform angular sampling. Figure 3(a) shows that α=0.60\alpha= 0.60 and 0.61 achieve the highest mean accuracy, with α=0.60\alpha= 0.60 performing better under CL and DN. We therefore adopt α=0.60\alpha= 0.60.

Figure 3. Parameter sensitivity analysis on CCPG. (a) Ablation study of the gait-oriented center parameter α\alpha. (b) Ablation study of the lower-body sampling ratio β\beta.

For LBB-NAS, we fix α=0.60\alpha= 0.60 and vary the proportion β\beta of angular samples assigned to the lower body. As shown in Figure 3(b), increasing β\beta from 0.50 to 0.75 improves the mean Rank-1 accuracy from 69.3% to 73.9%, with gains of 4.9%, 8.6%, and 5.9% under CL, UP, and BG, respectively. Further increasing β\beta reduces performance, particularly under DN. This suggests that denser lower-body sampling benefits gait representation, while excessive allocation weakens the balance between upper- and lower-body information. We therefore set β=0.75\beta= 0.75 by default.

Generalization and Robustness

Robustness Across Sketch Extractors

We investigate how different Sketch extractors affect the performance of PolarGait. As shown in Table V, TEED outperforms MUGE under clothing changes (57.4% vs. 49.7%), whereas MUGE performs better at nighttime (67.4% vs. 61.2%). Although MUGE achieves higher baseline OA@R-1 (92.1% vs. 91.3%), TEED gains more from FOSD (+0.6% vs. +0.2%) and reaches 93.8% with additional Silhouette, slightly exceeding MUGE (93.6%).

ConfigurationTEED NMTEED CLTEED NTTEED OA@R-1MUGE NMMUGE CLMUGE NTMUGE OA@R-1
Baseline95.257.461.291.395.549.767.492.1
+FOSD95.759.562.191.995.549.562.192.3
+FOSD+Silh.96.665.670.893.896.054.073.293.6

Table V. ABLATION RESULTS WITH DIFFERENT SKETCH EXTRACTORS ON SUSTech1K.

As illustrated in Figure 4, TEED preserves richer internal responses under clothing changes but produces more fragmented edges at nighttime, whereas MUGE maintains cleaner and more coherent contours. To quantify these differences, we measure edge density within body region RR and temporal intersection over union (T-IoU) between consecutive Sketch frames:

Qualitative comparison of TEED and MUGE on SUSTech1K

Figure 4. Qualitative comparison of TEED and MUGE on SUSTech1K under clothing change and nighttime conditions.

D(R)=∑t∣Et∩MtR∣∑t∣MtR∣(8)D(R) = \frac{\sum_t \lvert E_t \cap M_t^R \rvert}{\sum_t \lvert M_t^R \rvert} \tag*{(8)}
T-IoU=1T−1∑t=1T−1∣Et∩Et+1∣∣Et∪Et+1∣,(9)\mathrm{T\text{-}IoU} = \frac{1}{T-1}\sum_{t=1}^{T-1}\frac{\lvert E_t \cap E_{t+1} \rvert}{\lvert E_t \cup E_{t+1} \rvert}, \tag*{(9)}

where EtE_t denotes the binary Sketch response and MtRM_t^R denotes the mask of body region RR at frame tt. As shown in Figure 5, TEED exhibits higher edge density, particularly in the upper body under CL, but lower T-IoU under NT, indicating richer local responses with greater frame-to-frame variation.

Bar charts comparing overall edge density, upper-body edge density, lower-body edge density, and temporal IoU for MUGE and TEED under CL and NT conditions

Figure 5. Quantitative comparison of TEED and MUGE under clothing change (CL) and nighttime (NT) conditions. (a) Overall edge density. (b) Upper-body edge density. (c) Lower-body edge density. (d) Temporal intersection over union (IoU).

These characteristics help explain the different benefits of complementary representation learning. The denser responses of TEED provide more local variations for FOSD to reorganize, while its less explicit global contours benefit more from Silhouette fusion. In contrast, MUGE already preserves relatively coherent body structures, resulting in smaller additional gains. Overall, the effectiveness of PolarGait is influenced by the response characteristics of the input Sketch.

Cross-Domain Generalization

We further evaluate the cross-domain generalization of PolarGait between CCPG and SUSTech1K. As shown in Table VI, PolarGait consistently improves the corresponding Sketch baselines in both transfor directions. For CCPG→\toSUSTech1K, the MUGE variant improves CL, NT, and OA@R-1 from 22.9%, 27.0%, and 44.0% to 41.1%, 36.4%, and 56.0%, respectively, while the TEED variant also shows clear gains under CL and NT. For SUSTech1K→\toCCPG, PolarGait further improves the Sketch baseline under several conditions, particularly CL and UP. These results show that the complementary representations constructed by PolarGait improve cross-dataset transferability, especially under clothing variation and nighttime conditions.

InputMethodCCPG → SUSTech1K CLCCPG → SUSTech1K NTCCPG → SUSTech1K OA@R-1SUSTech1K → CCPG CLSUSTech1K → CCPG UPSUSTech1K → CCPG DNSUSTech1K → CCPG BG
Silh.DeepGaitV2 [31]17.614.723.013.017.926.839.4
ParsingDeepGaitV2 [31]24.322.646.311.017.732.860.0
Sketch (TEED)DeepGaitV2 [31]20.720.335.08.114.622.548.9
Sketch (MUGE)DeepGaitV2 [31]22.927.044.09.320.521.050.5
RGBBigGait [3]43.744.856.44.511.511.945.5
RGB+Silh.Gait-X [21]47.943.473.47.215.122.256.0
Sketch+Silh.+Pars.SketchGait++ (TEED) [17]30.024.651.512.322.134.270.3
Sketch+Silh.PolarGait (TEED)37.635.053.610.820.426.257.7
Sketch+Silh.PolarGait (MUGE)41.136.456.011.424.522.756.5

Table VI. CROSS-DATASET GENERALIZATION RESULTS. MODELS ARE TRAINED ON THE SOURCE DATASET AND EVALUATED ON THE TARGET DATASET.

Qualitative Visualization

We visualize the FOSD representations and their corresponding Grad-CAM responses across a representative gait sequence. As shown in Figure 6, the smoothed and residual branches exhibit distinct spatial activation patterns across gait frames, indicating different discriminative emphasis despite originating from the same Sketch. These observations provide qualitative support for the complementary feature learning encouraged by FOSD.

Qualitative visualization of FOSD across a representative gait sequence

Figure 6. Qualitative visualization of FOSD.

Limitations and Future Directions

PolarGait still has several limitations. First, the effectiveness of FOSD depends on the response richness of the input Sketch, and its benefit may be reduced when the representation is already sparse or structurally simplified. Second, GOPT relies on predefined geometric choices, including the polar center and the lower-body sampling ratio, which may not be optimal across different datasets or walking conditions. In addition, the current study focuses on Human Sketch as the primary representation for complementarity construction. Future work could explore adaptive filtering and spatial reparameterization, as well as whether similar principles can be extended to other gait representations.

Conclusion

In this work, we proposed PolarGait to explore whether complementary gait information can be actively constructed from a single representation, rather than relying solely on heterogeneous inputs. Using Human Sketch as the primary representation, FOSD constructs complementary response views through Gaussian smoothing and absolute residual extraction, while GOPT reorganizes structural information through gait-oriented polar reparameterization. Silhouette further provides complementary global body-shape information. Extensive experiments demonstrate the effectiveness of these constructed views across diverse gait recognition benchmarks. These findings suggest that complementary gait information can arise not only from heterogeneous representations, but also from different response patterns and spatial organizations within a single structured representation. Future work will investigate adaptive view construction and extend this principle to other gait representations.

Acknowledgements

This work was supported by the National Natural Science Foundation of China under Grant Nos. 62522308, 62673078, 62325302, and T2421001, the Natural Science Foundation of Chongqing, China under Grant No. CSTB2024NSCQ-MSX0472, and the Fundamental Research Funds for the Central Universities under Grant Nos. 2024CDJCGJ-006, 2025CDJZKKYJH-06, and 2025CDJGF-002.

References

References

  1. [1]I. Bouchrika, A survey of using biometrics for smart visual surveillance: Gait recognition, in: Surveillance in Action, Springer, 2017, 3–23.DOI
  2. [2]D. Jin, C. Fan, J. Ma, J. Zhou, W. Chen, S. Yu, On denoising walking videos for gait recognition, in: Proc. IEEE/CVF CVPR, 2025, 12347–12357.
  3. [3]D. Ye, C. Fan, J. Ma, X. Liu, S. Yu, BigGait: Learning gait representation you want by large vision models, in: Proc. IEEE/CVF CVPR, 2024, 200–210.
  4. [4]H. Chao, Y. He, J. Zhang, J. Feng, GaitSet: Regarding gait as a set for cross-view gait recognition, in: Proc. AAAI, 33, 2019, 8126–8133.
  5. [5]C. Fan, Y. Peng, C. Cao, X. Liu, S. Hou, J. Chu, Y. Huang, Q. Li, Z. He, GaitPart: Temporal part-based model for gait recognition, in: Proc. IEEE/CVF CVPR, 2020, 14225–14233.
  6. [6]B. Lin, S. Zhang, X. Yu, Gait recognition via effective global-local feature representation and local temporal aggregation, in: Proc. IEEE/CVF ICCV, 2021, 14648–14656.
  7. [7]J. Zheng, X. Liu, S. Wang, L. Wang, C. Yan, W. Liu, Parsing is all you need for accurate gait recognition in the wild, in: Proc. ACM MM, 2023, 116–124.
  8. [8]Z. Wang, S. Hou, M. Zhang, X. Liu, C. Cao, Y. Huang, GaitParsing: Human semantic parsing for gait recognition, IEEE Trans. Multimed. 26 (2024) 4736–4748. doi:10.1109/TMM.2023.3325962.
  9. [9]T. Teepe, A. Khan, J. Gilg, F. Herzog, S. Hörmann, G. Rigoll, GaitGraph: Graph convolutional network for skeleton-based gait recognition, in: Proc. IEEE ICIP, 2021, 2314–2318.
  10. [10]C. Fan, S. Hou, J. Liang, C. Shen, J. Ma, D. Jin, Y. Huang, S. Yu, OpenGait: A comprehensive benchmark study for gait recognition toward better practicality, IEEE Trans. Pattern Anal. Mach. Intell. 47 (10) (2025) 8397–8414. doi:10.1109/TPAMI.2025.3576283.
  11. [11]Y. Peng, K. Ma, Y. Zhang, Z. He, Learning rich features for gait recognition by integrating skeletons and silhouettes, Multimed. Tools Appl. 83 (3) (2024) 7273–7294.
  12. [12]Y. Cui, Y. Kang, Multi-modal gait recognition via effective spatial-temporal feature fusion, in: Proc. IEEE/CVF CVPR, 2023, 17949–17957.
  13. [13]Y. Sun, X. Feng, X. Liu, L. Ma, L. Hu, M. S. Nixon, TriGait: Hybrid fusion strategy for multimodal alignment and integration in gait recognition, IEEE Trans. Biom. Behav. Identity Sci. 7 (1) (2025) 82–94. doi:10.1109/TBIOM.2024.3435046.
  14. [14]J. Zheng, X. Liu, B. Zhang, C. Yan, J. Zhang, W. Liu, Y. Zhang, It takes two: Accurate gait recognition in the wild via cross-granularity alignment, in: Proc. ACM MM, 2024, 8786–8794.
  15. [15]D. Jin, C. Fan, W. Chen, S. Yu, Exploring more from multiple gait modalities for human identification, in: Proc. AAAI, 39, 2025, 4120–4128.
  16. [16]C. Zhang, Z. Mei, M. Wang, L. Dong, TriFusionGait: A unified framework for robust trimodal gait recognition, IEEE Sens. J. (2026). doi:10.1109/JSEN.2026.3687088.
  17. [17]C. Zhang, Z. Zheng, R. Li, and Z. Mei, Sketch it out: Exploring label-free structural cues for multimodal gait recognition, arXiv preprint arXiv:2603.05537, 2026. (Accepted by ACM Multimedia 2026.)
  18. [18]Q. Zhou, H. Zou, G. Wu, C. Liang, Z. Han, Z. Wang, GaitFuse: A hierarchical cross-modal alignment and uncertainty-aware fusion framework for multi-modal gait recognition, Inf. Fusion 137 (2027) 104615. doi:10.1016/j.inffus.2026.104615.
  19. [19]Q. Yang, A. Wu, W.-S. Zheng, Person re-identification by contour sketch under moderate clothing change, IEEE Trans. Pattern Anal. Mach. Intell. 43 (6) (2021) 2029–2046. doi:10.1109/TPAMI.2019.2960509.
  20. [20]S. Zhang, J. Zheng, S. Zhu, C. Yan, TrackletGait: A robust framework for gait recognition in the wild, IEEE Trans. Multimed. 27 (2025) 8875–8887. doi:10.1109/TMM.2025.3607705.
  21. [21]Z. Wang, S. Hou, J. Li, X. Liu, C. Cao, Y. Huang, S. Wang, M. Zhang, Gait-X: Exploring X modality for generalized gait recognition, in: Proc. IEEE/CVF CVPR, 2025, 13259–13269.
  22. [22]T. Teepe, J. Gilg, F. Herzog, S. Hörmann, G. Rigoll, Towards a deeper understanding of skeleton-based gait recognition, in: Proc. IEEE/CVF CVPR Workshops, 2022, 1569–1577.
  23. [23]A. Catruna, A. Cosma, E. Radoi, GaitPT: Skeletons are all you need for gait recognition, in: Proc. IEEE FG, 2024, 1–10.
  24. [24]M. Jaderberg, K. Simonyan, A. Zisserman, K. Kavukcuoglu, Spatial transformer networks, in: Adv. Neural Inf. Process. Syst., 28, 2015.arxiv.org/abs/1506.02025
  25. [25]J. Dai, H. Qi, Y. Xiong, Y. Li, G. Zhang, H. Hu, Y. Wei, Deformable convolutional networks, in: Proc. IEEE ICCV, 2017, 764–773.arxiv.org/abs/1703.06211
  26. [26]C. Esteves, C. Allen-Blanchette, X. Zhou, K. Daniilidis, Polar transformer networks, in: Proc. ICLR, 2018.arxiv.org/abs/1709.01889
  27. [27]B. Khaleghi, A. Khamis, F. O. Karray, S. N. Razavi, Multisensor data fusion: A review of the state-of-the-art, Inf. Fusion 14 (1) (2013) 28–44. doi:10.1016/j.inffus.2011.08.001.
  28. [29]F. Min, S. Guo, F. Hao, J. Dong, GaitMA: Pose-guided multi-modal feature fusion for gait recognition, in: Proc. IEEE ICME, 2024, 1–6.arxiv.org/abs/2407.14812
  29. [30]Z. Wu, C. Zhang, H. Xu, P. Jiao, H. Wang, DAGait: Generalized skeleton-guided data alignment for gait recognition, arXiv:2503.18830, 2025.arxiv.org/abs/2503.18830
  30. [31]C. Fan, S. Hou, Y. Huang, S. Yu, Exploring deep models for practical gait recognition, arXiv:2303.03301, 2023.arxiv.org/abs/2303.03301
  31. [32]C. Fan, J. Liang, C. Shen, S. Hou, Y. Huang, S. Yu, OpenGait: Revisiting gait recognition toward better practicality, in: Proc. IEEE/CVF CVPR, 2023, 9707–9716.arxiv.org/abs/2211.06597
  32. [33]H. Luo, Y. Gu, X. Liao, S. Lai, W. Jiang, Bag of tricks and a strong baseline for deep person re-identification, in: Proc. IEEE/CVF CVPR Workshops, 2019.arxiv.org/abs/1903.07071
  33. [34]A. Goyal, H. Law, B. Liu, A. Newell, J. Deng, Revisiting point cloud shape classification with a simple and effective baseline, in: Proc. ICML, PMLR, 2021, 3809–3820.arxiv.org/abs/2106.05304
  34. [35]C. Shen, C. Fan, W. Wu, R. Wang, G. Q. Huang, S. Yu, LiDAR-Gait: Benchmarking 3D gait recognition with point clouds, in: Proc. IEEE/CVF CVPR, 2023, 1054–1063.arxiv.org/abs/2211.10598
  35. [36]X. Han, Y. Ren, P. Cong, Y. Sun, J. Wang, L. Xu, Y. Ma, Gait recognition in large-scale free environment via single LiDAR, in: Proc. ACM MM, 2024, 380–389.arxiv.org/abs/2211.12371
  36. [37]C. R. Qi, L. Yi, H. Su, L. J. Guibas, PointNet++: Deep hierarchical feature learning on point sets in a metric space, in: Adv. Neural Inf. Process. Syst., 30, 2017.arxiv.org/abs/1706.02413
  37. [38]C. Shen, R. Wang, L. Duan, S. Yu, LidarGait++: Learning local features and size awareness from LiDAR point clouds for 3D gait recognition, in: Proc. IEEE/CVF CVPR, 2025, 6627–6636.DOI
  38. [39]Q. Cai, S. Hou, X. Hu, Y. Huang, BarbieGait: An identity-consistent synthetic human dataset with versatile cloth-changing for gait recognition, in: Proc. IEEE/CVF CVPR, 2026.arxiv.org/abs/2604.12221
  39. [40]H. Xu, Z. Wu, C. Zhang, Z. Chen, Z. Liu, P. Jiao, H. Wang, PSGait: Gait recognition using parsing skeleton, arXiv:2503.10247v3, 2026.
  40. [41]A. Hermans, L. Beyer, B. Leibe, In defense of the triplet loss for person re-identification, arXiv:1703.07737, 2017.arxiv.org/abs/1703.07737
  41. [42]W. Li, S. Hou, C. Zhang, C. Cao, X. Liu, Y. Huang, Y. Zhao, An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions, in: Proc. IEEE/CVF CVPR, 2023, 13824–13833.DOI
  42. [43]S. Yu, D. Tan, T. Tan, A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition, in: Proc. ICPR, 4, 2006, 441–444.DOI
  43. [44]J. Zheng, X. Liu, W. Liu, L. He, C. Yan, T. Mei, Gait recognition in the wild with dense 3D representations and a benchmark, in: Proc. IEEE/CVF CVPR, 2022, 20228–20237.arxiv.org/abs/2204.02569
  44. [45]X. Soria, Y. Li, M. Rouhani, A. D. Sappa, Tiny and efficient model for the edge detection generalization, in: Proc. IEEE/CVF ICCV Workshops, 2023, 1364–1373.arxiv.org/abs/2308.06468
  45. [46]C. Zhou, Y. Huang, M. Pu, Q. Guan, R. Deng, H. Ling, MuGE: Multiple granularity edge detection, in: Proc. IEEE/CVF CVPR, 2024, 25952–25962.DOI
  46. [47]N. Carion, L. Gustafson, Y.-T. Hu, et al., SAM 3: Segment anything with concepts, arXiv:2511.16719, 2025.
  47. [48]D. Ye, C. Fan, Z. Huang, C. Luo, J. Li, S. Yu, X. Liu, BiggerGait: Unlocking gait recognition with layer-wise representations from large vision models, in: Adv. Neural Inf. Process. Syst. 38, 2025.arxiv.org/abs/2505.18132
  48. [49]C. Zhang, X.-P. Chen, G.-Q. Han, X.-J. Liu, Spatial transformer network on skeleton-based gait recognition, Expert Syst. (2023) e13244. doi:10.1111/exsy.13244.DOI
  49. [50]Y. Jeon, J. Kim, Active convolution: Learning the shape of convolution for image classification, in: Proc. IEEE CVPR, 2017, 4201–4209.arxiv.org/abs/1703.09076
  50. [51]X. Wang, L. Xie, C. Dong, and Y. Shan, Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data, in: Proc. ICCV Workshops, 2021, pp. 1905–1914.arxiv.org/abs/2107.10833

Paper details

vlnc:W_cukcxxg5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents