Gamba：面向动作识别的基于 Mamba 的动态图拓扑学习图卷积网络
Gamba: Mamba-based graph convolutional network with dynamic graph topology learning for action recognition
用 Mamba 状态空间模型替代自注意力来学习骨架序列的动态邻接矩阵，为视觉中需要建模长程关节/节点关系的时序任务提供了一种计算复杂度更低的图拓扑学习范式。
原文依据 the state space model is introduced to process enhanced temporal features, aiming to learn dynamic adjacency matrices for long-range dependencies
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论现有图模型大多利用自注意力机制建模每个样本各关节之间的特征相关性，这不仅忽略了时序维度上的动态关系依赖，还导致冗余计算，并且难以建立统一的关节关系表示框架。为解决这些问题，本文提出了一种基于 Mamba 的图卷积网络（Gamba），并引入动态图拓扑学习。为捕捉局部运动模式，本文设计了一个节点分类模块，将运动关节划分为不同类型。据我们所知，这是首个为运动关节赋予标签信息以促进相关性学习的工作。为捕捉不同类别关节之间的潜在关系，本文引入状态空间模型处理增强后的时序特征，旨在为不同类别关节之间的长程依赖学习动态邻接矩阵。所提出的框架不仅有助于对时空特征建模进行自适应关注，而且相比传统基于自注意力的方法具有更低的计算复杂度。在公开的 NTU RGB+D 60/120 与 NW-UCLA 基准数据集上的大量实验表明，Gamba 在识别精度上优于最优方法。Gamba 的源代码可在 https://github.com/RCEricZhou/Gamba 获取。
英文原文摘要 1453 字符
Existing graph models predominantly utilize self-attention mechanisms to model feature correlations between the joints of each sample, which not only neglects dynamic relation dependencies in temporal dimension but also leads to redundant computation and difficulty in establishing a unified framework for joint relation representation. To address these problems, this paper develops a Mamba-based graph convolution network (Gamba) with dynamic graph topology learning. In order to capture local motion patterns, a node classification module has been developed to categorize motion joints into distinct types. To the best of our knowledge, this is the first work to assign motion joints with label information to facilitate correlation learning. To capture the underlying relation of the joints of different categories, the state space model is introduced to process enhanced temporal features, aiming to learn dynamic adjacency matrices for long-range dependencies of the joints across different categories. The proposed framework not only facilitates an adaptive focus on spatio-temporal feature modeling but also has less computational complexity than traditional self-attention-based approaches. Extensive experiments on the public NTU RGB+D 60/120 and NW-UCLA benchmark datasets demonstrate the superiority of Gamba over state-of-the-art methods in recognition accuracy. The source code of Gamba is available at https://github.com/RCEricZhou/Gamba.