Gaitmixer: Skeleton-Based Gait Representation Learning Via Wide-Spectrum Multi-Axial Mixer

A Gait Recognition Method via Graph Convolutional Networks With Local Mask

“…In this section, we constructed the final MA‐Gait model using the local mask $$$ {M}_P $$$ and global mask $$$ {M}_G $$$. Subsequently, the comparation experiments were conducted with several state‐of‐the‐art gait recognition methods on the CASIA‐B dataset [15], including PoseGait [6], GaitGraph [7], GaitGraph2 [9], LUGAN [10], CycleGait [11], and GaitMixer [8]. The results are presented in Table 3.…”

“…Following normalization according to Ref. [8], the original joint sequence $$$ {\boldsymbol{H}}_{\boldsymbol{o}} $$$ was obtained in Equation (), where $$$ {v}_{t,n}\left({x}_{t,n},{y}_{t,n}\right) $$$ represents the 2D coordinates of the pedestrian's $$$ n $$$th joint in the $$$ t $$$th frame. $$$$ {\boldsymbol{H}}_{\boldsymbol{o}}=\left\{{v}_{t,n}\left({x}_{t,n},{y}_{t,n}\right)|t<T,n=0,1,\dots, 16\right\} $$$$ …”

“…Specifically, it outperforms CycleGait [11] by 5.0% in BG and 11.3% in CL. Secondly, compared with transformer‐based method of GaitMixer [8] at the viewing angle of 90°, MA‐Gait shows slightly lower performance in NM (by 1.3%) and BG (by 1.8%). This performance reduction is due to the weakened coordination movements between arms and legs in the 2D joint sequence at a 90° viewing.…”

“…However, it achieves a significant 5.2% improvement in the challenging CL condition. In terms of average cross‐walking condition accuracy, MA‐Gait (91.37%) surpasses GaitMixer [8] (88.33%) by 3.03%. These improvements are attributed to the masking mechanism, which guides the MA‐Gait to focus on the motion patterns of key body parts (arms, legs, head, and trunk), thereby enhancing robustness to resist carrying and clothing variations.…”

A Gait Recognition Method via Graph Convolutional Networks With Local Mask

“…In this section, we constructed the final MA‐Gait model using the local mask $$$ {M}_P $$$ and global mask $$$ {M}_G $$$. Subsequently, the comparation experiments were conducted with several state‐of‐the‐art gait recognition methods on the CASIA‐B dataset [15], including PoseGait [6], GaitGraph [7], GaitGraph2 [9], LUGAN [10], CycleGait [11], and GaitMixer [8]. The results are presented in Table 3.…”

“…Following normalization according to Ref. [8], the original joint sequence $$$ {\boldsymbol{H}}_{\boldsymbol{o}} $$$ was obtained in Equation (), where $$$ {v}_{t,n}\left({x}_{t,n},{y}_{t,n}\right) $$$ represents the 2D coordinates of the pedestrian's $$$ n $$$th joint in the $$$ t $$$th frame. $$$$ {\boldsymbol{H}}_{\boldsymbol{o}}=\left\{{v}_{t,n}\left({x}_{t,n},{y}_{t,n}\right)|t<T,n=0,1,\dots, 16\right\} $$$$ …”

“…Specifically, it outperforms CycleGait [11] by 5.0% in BG and 11.3% in CL. Secondly, compared with transformer‐based method of GaitMixer [8] at the viewing angle of 90°, MA‐Gait shows slightly lower performance in NM (by 1.3%) and BG (by 1.8%). This performance reduction is due to the weakened coordination movements between arms and legs in the 2D joint sequence at a 90° viewing.…”

“…However, it achieves a significant 5.2% improvement in the challenging CL condition. In terms of average cross‐walking condition accuracy, MA‐Gait (91.37%) surpasses GaitMixer [8] (88.33%) by 3.03%. These improvements are attributed to the masking mechanism, which guides the MA‐Gait to focus on the motion patterns of key body parts (arms, legs, head, and trunk), thereby enhancing robustness to resist carrying and clothing variations.…”

Explicit Time-Frequency Dynamics for Skeleton-Based Gait Recognition

“…Skeleton-based. Joint-graph methods including PoseGait [8], GaitGraph/GaitGraph2 [9, 10], and GaitMixer [11] model kinematic structure and are naturally robust to appearance changes. However, they depend on pose-estimation accuracy and typically underuse explicit motion dynamics.…”

A Gait Recognition Method via Graph Convolutional Networks With Local Mask

“…In this section, we constructed the final MA‐Gait model using the local mask $$$ {M}_P $$$ and global mask $$$ {M}_G $$$. Subsequently, the comparation experiments were conducted with several state‐of‐the‐art gait recognition methods on the CASIA‐B dataset [15], including PoseGait [6], GaitGraph [7], GaitGraph2 [9], LUGAN [10], CycleGait [11], and GaitMixer [8]. The results are presented in Table 3.…”

“…Following normalization according to Ref. [8], the original joint sequence $$$ {\boldsymbol{H}}_{\boldsymbol{o}} $$$ was obtained in Equation (), where $$$ {v}_{t,n}\left({x}_{t,n},{y}_{t,n}\right) $$$ represents the 2D coordinates of the pedestrian's $$$ n $$$th joint in the $$$ t $$$th frame. $$$$ {\boldsymbol{H}}_{\boldsymbol{o}}=\left\{{v}_{t,n}\left({x}_{t,n},{y}_{t,n}\right)|t<T,n=0,1,\dots, 16\right\} $$$$ …”

“…Specifically, it outperforms CycleGait [11] by 5.0% in BG and 11.3% in CL. Secondly, compared with transformer‐based method of GaitMixer [8] at the viewing angle of 90°, MA‐Gait shows slightly lower performance in NM (by 1.3%) and BG (by 1.8%). This performance reduction is due to the weakened coordination movements between arms and legs in the 2D joint sequence at a 90° viewing.…”

“…However, it achieves a significant 5.2% improvement in the challenging CL condition. In terms of average cross‐walking condition accuracy, MA‐Gait (91.37%) surpasses GaitMixer [8] (88.33%) by 3.03%. These improvements are attributed to the masking mechanism, which guides the MA‐Gait to focus on the motion patterns of key body parts (arms, legs, head, and trunk), thereby enhancing robustness to resist carrying and clothing variations.…”

Explicit Time-Frequency Dynamics for Skeleton-Based Gait Recognition

“…Skeleton-based. Joint-graph methods including PoseGait [8], GaitGraph/GaitGraph2 [9, 10], and GaitMixer [11] model kinematic structure and are naturally robust to appearance changes. However, they depend on pose-estimation accuracy and typically underuse explicit motion dynamics.…”

A Gait Recognition Method via Graph Convolutional Networks With Local Mask

“…In this section, we constructed the final MA‐Gait model using the local mask $$$ {M}_P $$$ and global mask $$$ {M}_G $$$. Subsequently, the comparation experiments were conducted with several state‐of‐the‐art gait recognition methods on the CASIA‐B dataset [15], including PoseGait [6], GaitGraph [7], GaitGraph2 [9], LUGAN [10], CycleGait [11], and GaitMixer [8]. The results are presented in Table 3.…”

“…Following normalization according to Ref. [8], the original joint sequence $$$ {\boldsymbol{H}}_{\boldsymbol{o}} $$$ was obtained in Equation (), where $$$ {v}_{t,n}\left({x}_{t,n},{y}_{t,n}\right) $$$ represents the 2D coordinates of the pedestrian's $$$ n $$$th joint in the $$$ t $$$th frame. $$$$ {\boldsymbol{H}}_{\boldsymbol{o}}=\left\{{v}_{t,n}\left({x}_{t,n},{y}_{t,n}\right)|t<T,n=0,1,\dots, 16\right\} $$$$ …”

“…Specifically, it outperforms CycleGait [11] by 5.0% in BG and 11.3% in CL. Secondly, compared with transformer‐based method of GaitMixer [8] at the viewing angle of 90°, MA‐Gait shows slightly lower performance in NM (by 1.3%) and BG (by 1.8%). This performance reduction is due to the weakened coordination movements between arms and legs in the 2D joint sequence at a 90° viewing.…”

“…However, it achieves a significant 5.2% improvement in the challenging CL condition. In terms of average cross‐walking condition accuracy, MA‐Gait (91.37%) surpasses GaitMixer [8] (88.33%) by 3.03%. These improvements are attributed to the masking mechanism, which guides the MA‐Gait to focus on the motion patterns of key body parts (arms, legs, head, and trunk), thereby enhancing robustness to resist carrying and clothing variations.…”

Explicit Time-Frequency Dynamics for Skeleton-Based Gait Recognition

“…Skeleton-based. Joint-graph methods including PoseGait [8], GaitGraph/GaitGraph2 [9, 10], and GaitMixer [11] model kinematic structure and are naturally robust to appearance changes. However, they depend on pose-estimation accuracy and typically underuse explicit motion dynamics.…”