Transformer-based models have excelled in 3D human pose estimation tasks by effectively encoding the interconnections among all joints, thus capturing the global dependencies inherent in the data. However, such methods often fail to adequately model the local dependencies of joints and suffer from a quadratic increase in computational complexity with sequence length, leading to a waste of computational resources. To address these issues, this study defines the human motion process as a sequence of input-output processes of a state-space model and proposes a 3D human pose estimation method, GMambaScanX, based on the fusion of GCN and Mamba. The GCN model captures the spatiotemporal local dependencies of human motion sequences, thereby enhancing the model's ability to capture local relationships. The Mamba model captures the spatiotemporal long-range dependencies of human motion sequences, thereby improving the model's ability to capture global relationships. We present two MambaScan scanning strategies: a sequential joint scan for temporal motion state transitions, which enhances the model's comprehension of human motion features within the temporal dimension, and a spatial joint scan informed by human structural prior knowledge, which augments the model's understanding of human structural features in the spatial domain. GMambaScanX was trained and validated on the Human3.6M dataset, achieving an average joint position error of 39.8 mm with only 12.4% (5.3×106) of the parameters of MotionBERT. Compared with fully connected models, GMambaScanX better extracts human structural features, utilizes parameters more efficiently, and improves the model's effectiveness and efficiency.
状态空间模型[5](State Space Model,SSM)是一种用于描述动态系统行为的数学模型,不同于Transformer采用注意力机制获取不同时序状态的全局依赖关系,SSM用一组一阶微分方程(连续时间系统)或差分方程(离散时间系统)表示系统的内部状态的演化,同时用另一组方程描述系统状态和输出之间的关系。这些方程可以表示为矩阵和向量的形式,将系统离散化后,使用离散卡尔曼滤波器对系统状态进行估计,实现网络的推理与训练。Gu等[6]在此基础上提出了选择性状态空间模型(Mamba)。Mamba模型是一种基于状态空间模型(SSM)的新型深度学习架构,能够在保持线性计算复杂度的同时,有效建模长程依赖关系,这一特性使其在处理长序列数据时表现出色,其架构设计独特,引入时变参数动态调整模型的状态,允许模型根据输入数据选择性地传递或遗忘信息,从而能够更有效的处理离散和长序列数据。
SMITHA C, BROWNE N. Estimating a state-space model from point process observations[J]. Neural Computation, 2003, 15(5): 965-991. DOI: 10.1162/089976603765202622 .
[6]
GUA, DAO T. Mamba: Linear-time sequence modeling with selective state spaces[EB/OL]. 2023: 2312.00752.
[7]
IONESCUC, PAPAVAD, OLARUV, et al. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2014, 36(7): 1325-1339. DOI: 10.1109/TPAMI.2013.248 .
[8]
DAO T, GUA. Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality[EB/OL]. 2024: 2405.21060.
[9]
ZHUL H, LIAOB C, ZHANGQ, et al. Vision Mamba: Efficient visual representation learning with bidirectional state space model[EB/OL]. 2024: 2401.09417.
[10]
HWANGS, LAHOTIA, DAO T, et al. Hydra: Bidirectional state space models through generalized matrix mixers[EB/OL]. 2024: 2407.09941.
[11]
CHENY X, ZHANGZ Q, YUANC F, et al. Channel-wise topology refinement graph convolution for skeleton-based action recognition[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). New York: IEEE Press, 2021: 13339-13348. DOI: 10.1109/ICCV48922.2021.01311 .
[12]
CI H, WANGC Y, MAX X, et al. Optimizing network structure for 3D human pose estimation[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV).New York: IEEE Press, 2019: 2262-2271. DOI: 10.1109/ICCV.2019.00235 .
[13]
ZOUZ M, TANGW. Modulated graph convolutional network for 3D human pose estimation[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). New York: IEEE Press, 2021: 11457-11467. DOI: 10.1109/ICCV48922.2021.01128 .
[14]
YUB X B, ZHANGZ, LIUY X, et al. GLA-GCN: Global-local adaptive graph convolutional network for 3D human pose estimation from monocular video[C]//2023 IEEE/CVF International Conference on Computer Vision (ICCV). New York: IEEE Press, 2023: 8784-8795. DOI: 10.1109/ICCV51070.2023.00810 .
[15]
MEHTAD, RHODINH, CASASD, et al. Monocular 3D human pose estimation in the wild using improved CNN supervision[C]//2017 International Conference on 3D Vision (3DV). New York: IEEE Press, 2017: 506-516. DOI: 10.1109/3DV.2017.00064 .
[16]
LIW H, LIUH, TANGH, et al. MHFormer: Multi-hypothesis transformer for 3D human pose estimation[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2022: 13137-13146. DOI: 10.1109/CVPR52688.2022.01280 .
[17]
ZHANGJ L, TUZ G, YANGJ Y, et al. MixSTE: Seq2seq mixed spatio-temporal encoder for 3D human pose estimation in video[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2022: 13222-13232. DOI: 10.1109/CVPR52688.2022.01288 .
[18]
TANGZ H, QIUZ F, HAOY B, et al. 3D human pose estimation with spatio-temporal criss-cross attention[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2023: 4790-4799. DOI: 10.1109/CVPR52729.2023.00464 .
[19]
ZHAOQ T, ZHENGC, LIUM Y, et al. PoseFormerV2: Exploring frequency domain for efficient and robust 3D human pose estimation[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2023: 8877-8886. DOI: 10.1109/CVPR52729.2023.00857 .
[20]
ZHENGC, ZHUS J, MENDIETAM, et al. 3D human pose estimation with spatial and temporal transformers[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). New York: IEEE Press, 2021: 11636-11645. DOI: 10.1109/ICCV48922.2021.01145 .
[21]
CHENH Y, HEJ Y, XIANGW M, et al. HDFormer: High-order directed transformer for 3D human pose estimation[EB/OL]. 2023: 2302.01825. DOI: 10.24963/ijcai.2023/65 .
[22]
MEHRABANS, ADELIV, TAATIB. MotionAGFormer: Enhancing 3D human pose estimation with a transformer-GCNFormer network[C]//2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). New York: IEEE Press, 2024: 6905-6915. DOI: 10.1109/WACV57701.2024.00677 .