{
  "paper_id": "a8a21cdc14d1d7c5927ee224879e1132",
  "paper_title": "Deep Kinematics Analysis for Monocular 3D Human Pose Estimation",
  "paper_authors": [
    "Jingwei Xu",
    "Zhenbo Yu",
    "Bingbing Ni",
    "Jiancheng Yang",
    "Xiaokang Yang",
    "Wenjun Zhang"
  ],
  "publication_date": "",
  "publication_venue": "2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)",
  "abstract": "For monocular 3D pose estimation conditioned on 2D detection, noisy/unreliable input is a key obstacle in this task. Simple structure constraints attempting to tackle this problem, e.g., symmetry loss and joint angle limit, could only provide marginal improvements and are commonly treated as auxiliary losses in previous researches. It still remains challenging to fully utilize human prior knowledge in this task. In this paper, we propose to address the above issue in a systematic view. Firstly, we show that optimizing the kinematics structure of noisy 2D inputs is critical to obtain accurate 3D estimations. Secondly, based on corrected 2D joints, we further explicitly decompose articulated motion with human topology, which leads to more compact 3D static structure easier for estimation. Finally, we propose a temporal module to refine 3D trajectories, which obtains more rational results. Above three steps are seamlessly integrated into deep neural models, which form a deep kinematics analysis pipeline concurrently considering the static/dynamic structure of 2D inputs and 3D outputs. Extensive experiments show that proposed framework achieves state-of-the-art performance on two widely used 3D human action datasets. Meanwhile, targeted ablation study shows that each former step is critical for the latter one to obtain promising results.",
  "keywords": [
    "3D human pose estimation",
    "kinematics analysis",
    "monocular video",
    "perspective projection",
    "pose decomposition",
    "trajectory completion"
  ],
  "paper_institutions": [
    "Shanghai Jiao Tong University",
    "MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University",
    "Huawei Hisilicon"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:47:36.253607Z",
  "template_tag": "simple",
  "summary": "### 摘要\n针对单目3D人体姿态估计中2D检测噪声大、不可靠的问题，现有方法多采用对称性损失、关节角度限制等简单结构约束，但效果有限且常作为辅助损失。本工作提出系统性利用运动学先验，构建深度运动学分析流水线。首先，基于透视投影约束优化2D关键点，纠正噪声并保持2D/3D对应关系；其次，利用人体刚性结构将3D姿态估计分解为骨骼长度与方向两个子任务，显著压缩输出空间并降低学习难度；最后，针对四肢等低置信度关节，通过轨迹补全的时序模块细化动态结构。三部分无缝集成到深度网络，同时考虑2D输入和3D输出的静态/动态结构。在Human3.6M和HumanEva两个广泛使用的数据集上进行实验，均达到最先进性能。消融实验表明每个模块对最终结果均有正向贡献，且前一步的优化对后续步骤至关重要。该方法适用于动作识别、人体重建和机器人操作等应用，但依赖固定相机参数，对于变焦场景需提供内参作为输入，这留作未来工作。核心机制是利用透视投影公式 a = x/z * f_x + c_x 及骨骼长度恒定假设，将坐标回归分解为长度和方向估计。\n### 关键要点\n- 2D检测噪声和不可靠是单目3D姿态估计的主要障碍，简单结构约束改善有限。\n- 提出基于透视投影的2D关键点优化方法，通过线性回归估计相机参数f和c，并施加L1损失和时域一致性约束。\n- 实验表明，仅用2D时序平滑（无投影约束）时训练/测试误差差异大，3D测试集改进微弱，投影约束有效缓解了该问题。\n- 3D姿态估计被分解为骨骼长度回归和方向回归两个子任务，输出空间维度从3*T*J降至2*T*J+J，显著降低学习难度。\n- 长度估计采用特征聚合与L1损失，方向估计用L2归一化与余弦相似度损失，最终通过骨架拓扑迭代组合得到完整3D姿态。\n- 轨迹补全模块针对四肢关节（低置信度）进行掩码丢弃，利用双向LSTM重构平滑轨迹，损失包含一阶和二阶时间差分项。\n- 在Human3.6M和HumanEva数据集上，以MPJPE和PA-MPJPE为指标，9帧输入下模型达到45.6mm和36.2mm误差，优于Pavllo等人（49.8mm/40.0mm）等对比方法。\n- 消融实验显示，3D分解模块贡献最大，2D投影修正和轨迹补全均带来一致精度提升。\n- 增加输入帧数可从1帧到9帧持续降低误差，模型在T=3、5、7、9时均优于相同输入长度的先前工作。\n- 定性可视化表明，对于严重遮挡动作（如坐下），模型能保持四肢结构和轨迹合理性，优于仅考虑时间平滑的方法。\n- 方法依赖固定相机内参，对变焦场景需提供(f,c)输入以保证泛化，这是当前局限性和未来研究方向。\n- 整体流水线可端到端训练，总损失包含2D时序损失、投影损失、3D分解损失、最终L2损失及补全损失，权重为1、0.1、0.1、1、1、0.1。"
}