DPAGait: Dynamic Pose-Knowledge-Aware Network for Robust Gait Recognition in the Wild
Authors: Zenan Lin, Junduan Huang, Kun Liu, M. Saad Shakeel, Wenxiong Kang
Organizations: School of Automation Science and Engineering, South China University of Technology, Guangzhou 510641, China · School of Artificial Intelligence, South China Normal University, Foshan 528225, China · Pazhou Laboratory, Guangzhou 510335, China · School of Future Technology, South China University of Technology, Guangzhou 510641, China
Gait biometrics aims at identity recognition by appearance or pose features captured during unconstrained human walking. Existing methods predominantly rely on binary silhouette images, inherently characterized by low information entropy, and fail to fully utilize the dynamic information in gait sequences, limiting their ability to handle cross-time variations effectively. Consequently, these methods encounter significant challenges in the wild. They are susceptible to variations in clothing, carried objects, and viewpoints, as well as dynamic environmental conditions, all of which lead to substantial degradation in recognition accuracy and robustness. To overcome these challenges, we propose a pose-knowledge-aware approach specifically designed for robust gait recognition in the wild. First, we propose a novel Dynamic Pose-Parsing enhanced Silhouette gait representation paradigm that substantially enhances the information entropy and effectively captures fine-grained behavioral cues by integrating pose-parsing cues with traditional silhouettes. Subsequently, we develop a dual-branch Dynamic Pose-Knowledge-Aware Gait network. The framework consists of two complementary branches: (1) a 2D dynamic pose-aware branch that extracts pose-enhanced features from our proposed DPPS representations, and (2) a 3D Skinned Multi-Person Linear branch that leverages the 3D SMPL model to capture rich and comprehensive 3D dynamic pose information. Finally, the features are integrated to facilitate accurate and robust gait recognition. Extensive experiments on four challenging in-the-wild gait datasets demonstrate the effectiveness and generalizability of the proposed DPPS representation. The complete framework is further evaluated on SUSTech1K, CCPG, and Gait3D, where SMPL information is available or can be consistently estimated, achieving competitive performance under diverse real-world conditions.
Figures & tables
Aspect
Sils
Pose
SMPL
Our DPPS
Representation
Modality
Image
Point/Graph
3D Mesh
Image
Body Shape
✓
✗
✓
✓
Structural
Silhouette Contour
Sparse Joints
Dense Mesh Vertices
Fine-grained Contour & Parts
Discriminability
Medium
Low
High
High
Size
64×64
17×2
85D / 6890×3
64×64
Table 1: Comparison of different gait representations.
Figure 1: Visualization of DPPS patch semantics. The highlighted region in the silhouette exhibits homogeneous pixel values (e.g., all 255), producing a nearly constant patch embedding (unit vector), thus carrying minimal discriminative information. In contrast, our DPPS preserves structural variations within the same region, yielding more informative dense vectors. This visualization highlights the fine-grained spatial structure encoded in our DPPS gait representation. Conv3 × 3 denotes a 3×3 convolution kernel.
Figure 2: Overview of the proposed DPAGait network. The network consists of two parallel branches: the 2D dynamic pose-aware branch takes DPPS as input to extract 2D pose-aware features, while the 3D SMPL branch leverages SMPL parameters to model 3D pose-aware behavioral representations. The two branches are fused using the Residual Attention Fusion (RAFusion) module for final gait recognition through joint optimization with reconstruction loss, triplet loss, and cross-entropy loss.
Figure 3: The proposed Residual Attention Fusion (RAFusion) module.
Modality
Method
SUSTech1K
CCPG
Probe Sequence (R-1)
Overall
CL
UP
DN
BG
Mean
NM
BG
CL
CR
UB
UN
OC
NT
R-1
R-5
R-1
mAP
R-1
mAP
R-1
mAP
R-1
mAP
R-1
Skeleton
GaitGraph2 [ 37 ]
22.2
18.2
6.8
18.6
13.4
19.2
27.3
16.4
18.6
40.2
5.0
2.5
5.3
4.0
5.8
4.2
6.2
4.6
5.6
GaitTR [ 48 ]
33.3
31.5
21.0
30.4
22.7
34.6
44.9
23.5
30.8
56.0
15.7
9.7
18.3
16.1
18.5
16.4
17.5
15.3
17.5
SkeletonGait [ 7 ]
55.0
51.0
24.7
49.9
42.3
52.0
62.8
43.9
50.1
72.6
40.4
20.8
48.5
35.8
53.0
40.3
61.7
48.1
50.9
Sils
GaitSet [ 3 ]
69.1
68.2
37.4
65.0
63.1
61.0
67.2
23.0
65.0
84.8
60.2
47.4
65.2
63.2
65.1
61.9
68.5
65.8
64.8
Table 2: Evaluation with different attributes on SUSTech1K and CCPG datasets.
Modality
Method
GREW
Gait3D
R-1
R-5
R-10
R-1
R-5
mAP
mINP
Sils
GaitSet [ 3 ]
46.3
63.6
70.3
36.7
58.3
30.0
17.3
GaitPart [ 8 ]
44.0
60.7
67.3
28.2
47.6
21.6
12.4
GaitGL [ 23 ]
47.3
63.6
69.3
29.7
48.5
22.3
13.6
GaitBase [ 6 ]
60.1
75.5
80.4
64.6
-
-
-
QAGait [ 44 ]
59.1
74.0
79.2
67.0
81.5
56.5
-
Table 3: Comparative results on two outdoor in-the-wild gait datasets, GREW and Gait3D.
Modality
Method
Backbone
GFLOPs
Parsing
SCHP
A-CE2P
21.9
Flow
VideoFlow
ResNet-like
21.0
DPPS
HRNet + DI
HRNet [ 41 ]
9.5
Table 4: GFLOPs comparison of different modalities at 256×256 resolution. The calculation protocol follows MultiGait++ [ 14 ] . “DI” denotes the Dynamic Image generation method [ 1 ] .
Figure 4: Comparison of Rank-1 accuracy between DPPS and silhouette inputs across representative single-stream gait recognition models on four benchmark datasets.
Model
Train on SUSTech1K
Train on Gait3D
Train on CCPG
CCPG
Gait3D
SUSTech1K
CCPG
SUSTech1K
Gait3D
CL
BG
Overall
OC
Overall
CL
BG
OC
Overall
Overall
SMPLGait [ 50 ]
14.7
41.0
12.9
35.7
33.8
11.1
59.5
10.9
11.5
8.6
DPGait [ 15 ]
7.9
28.7
–
–
–
–
–
18.9
14.3
–
DPAGait (Ours)
15.0
52.1
26.3
56.4
47.3
15.0
83.8
32.1
31.4
28.8
Table 5: Cross-dataset performance comparison among SUSTech1K, Gait3D, and CCPG. All models are trained only on the source dataset and directly evaluated on the target dataset without fine-tuning or domain adaptation. Results are reported in Rank-1 accuracy (%).
Method
Temporal Window Length
SUSTech1K
CCPG
Gait3D
PPS
–
81.0
85.7
79.2
RankSVM - Based Dynamic Encoding (DPPS)
2
81.2
85.7
79.6
3
81.3
85.9
81.1
4
80.7
85.3
80.0
Table 6: Ablation study of static pose - silhouette fusion and RankSVM - based dynamic encoding with different temporal window length n on SUSTech1K, CCPG, and Gait3D. The stride is fixed at s=1 for all dynamic encoding settings. Results are reported in Rank - 1 accuracy (%).
Method
2D Pose-Aware Branch
3D SMPL Branch
SUSTech1K
CCPG
Gait3D
Sche.1
Sils
×
77.4
83.3
74.4
Sche.2
Skeleton
×
47.3
60.5
22.4
Sche.3
PPS
×
81.0
85.7
79.2
Sche.4
DPPS
×
81.3
85.9
81.1
Sche.5(Ours)
DPPS
SMPL
85.6
86.4
81.6
Table 7: Ablation studies of the dual-branch architecture on three datasets under Rank-1 accuracy.
Method
SUSTech1K
CCPG
Gait3D
Mean
85.2
86.0
79.8
Concatenation
85.2
85.5
80.5
SMPLGait-style Fusion [ 50 ]
85.3
84.6
79.1
Attention [ 7 ]
83.6
82.1
80.3
RAFusion(Ours)
85.6
86.4
81.6
Table 8: Comparison of different fusion strategies on three datasets under Rank-1 accuracy.
Figure 5: The heatmaps generated by DeepGaitV2, Attention Fusion, and our RAFusion methods (from the second to the last row) on three datasets under different conditions: Normal, Night, and Occlusion.