{
  "paper_id": "81435aea3893405c39ca350cd811fb1b",
  "paper_title": "3D Human Pose Estimation from a Single Image via Distance Matrix Regression",
  "paper_authors": [
    "Francesc Moreno-Noguer"
  ],
  "publication_date": "",
  "publication_venue": "2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)",
  "abstract": "This paper addresses the problem of 3D human pose estimation from a single image. We follow a standard two-step pipeline by first detecting the 2D position of the N body joints, and then using these observations to infer 3D pose. For the first step, we use a recent CNN-based detector. For the second step, most existing approaches perform 2N-to-3N regression of the Cartesian joint coordinates. We show that more precise pose estimates can be obtained by representing both the 2D and 3D human poses using $N \\times N$ distance matrices, and formulating the problem as a 2D-to-3D distance matrix regression. For learning such a regressor we leverage on simple Neural Network architectures, which by construction, enforce positivity and symmetry of the predicted matrices. The approach has also the advantage to naturally handle missing observations and allowing to hypothesize the position of non-observed joints. Quantitative results on Humaneva and Human3.6M datasets demonstrate consistent performance gains over state-of-the-art. Qualitative evaluation on the images in-the-wild of the LSP dataset, using the regressor learned on Human3.6M, reveals very promising generalization results.",
  "keywords": [
    "3D human pose estimation",
    "distance matrix regression",
    "neural networks",
    "single image",
    "Euclidean distance matrices",
    "CNN",
    "occlusion handling"
  ],
  "paper_institutions": [
    "Institut de Robòtica i Informàtica Industrial (CSIC-UPC)"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:47:05.824194Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文针对单幅图像中三维人体姿态估计的严重病态问题，采用先检测二维关节再推断三维姿态的标准两阶段流程。所提出的方法将二维和三维人体姿态表示为关节间的欧氏距离矩阵（EDM），将问题形式化为二维到三维距离矩阵回归，从而自然地编码姿态结构信息并捕捉关节间相关性，减少了二维至三维映射的固有歧义。该方法使用简单的全连接网络（FConn）和全卷积网络（FConv）进行回归，通过构造保证输出矩阵的对称性和正性；对于遮挡关节，通过将对应行列置零实现自然处理。在Humaneva-I和Human3.6M数据集上进行了定量评估，在多种协议下均显著优于当时的最先进方法；同时在LSP自然图像上的定性结果展示了良好的泛化能力。模型对二维检测噪声和关节遮挡具有鲁棒性，其中FConv能合理推断遮挡肢体。核心机制是将二维EDM输入网络回归三维EDM，再通过多维缩放（MDS）恢复关节位置，并用基于人类形态学先验的反射消歧。\n### 关键要点\n- 使用欧氏距离矩阵（EDM）表示二维和三维姿态，并通过二维到三维EDM回归进行姿态估计，取代传统的笛卡尔坐标回归。\n- EDM表示天然包含姿态结构信息，且对旋转、平移和反射具有不变性，降低了二维到三维映射的歧义。\n- 采用两种神经网络架构：带128-128-91神经元的三层全连接网络（约40K参数）和带收缩-扩张结构的全卷积网络（约606K参数），均通过ReLU确保输出非负。\n- 通过矩阵对称化层和ReLU确保FConv输出满足EDM的对称性和正性，全连接网络直接回归上三角元素而保证对称。\n- 训练时使用L2损失和Adam优化器，FConn约500轮收敛，FConv约1500轮；在Humaneva-I上批量大小为7，在Human3.6M上为200。\n- 在Humaneva-I上，全连接网络和全卷积网络在Walking、Jogging、Boxing动作上的平均关节误差分别达到约19.20mm和19.20mm（GT+CPM训练），大幅优于先前方法。\n- 在Human3.6M上，采用FConv在协议1、2、3下的平均误差分别为85.64mm、81.52mm和73.98mm，多数情况下优于或接近现有最佳结果。\n- 通过合成关节遮挡（随机两关节或整肢体）训练网络，FConv在Humaneva-I上平均误差约32.68mm，对遮挡关节的误差约36.32mm，能较好推测遮挡肢体。\n- 在Human3.6M上随机双关节遮挡的平均误差为106.79mm，遮挡关节误差为118.95mm；特定动作如Sit、Photo的遮挡推测误差较大，源于训练样本稀少。\n- 对二维高斯噪声的鲁棒性测试显示，即使噪声标准差达20像素，三维误差仍可接受；而CPM检测的平均二维误差约10.91像素。\n- 在LSP数据集上，使用Human3.6M训练的网络进行跨域测试，通过PnP重投影的平均误差为8.83像素（基于CPM检测），表明了良好的泛化能力。\n- 由EDM恢复三维姿态需解决反射歧义，通过基于人体关节角度先验的形态学判断选择正确的镜像。"
}