End-to-End Feature Pyramid Network for Real-Time Multi-Person Pose Estimation
End-to-End Feature Pyramid Network for Real-Time Multi-Person Pose Estimation
复制标题
用于实时多人姿势估计的端到端特征金字塔网络
DOI：
10.23919/mva.2019.8758029
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
T. Ikenaga
中科院分区：
文献类型：
--
作者：
Dingli Luo;Songlin Du;T. Ikenaga
在计算机视觉中，姿态估计系统被广泛用于构造人体变换。然而，要同时实现稳定的实时速度、多变的人员数量和较高的准确率，是一件很困难的事情。本文提出了一种端到端的姿态估计网络。它包含一个神经网络友好的人类姿势表示。然后提出了一种基于特征金字塔网络结构的端到端实时姿态估计网络，该网络包含了基于注意力的检测模块。该网络可以在GTX 1070上以384 x 384的分辨率以超过60 fps的速度检测多个人，并且具有可负担的准确性。这项工作表明，与最先进的结果相比，这种网络结构的潜力可以更快，更好地执行。
In computer vision, pose estimation system is widely used to construct human body transformation. However, it is hard to achieve these targets together: stable real-time speed, variance human number and high accuracy. This paper proposes an end-to-end pose estimation network. It contains a neural network friendly representation of human pose. Then it proposes a correspond real-time end-to-end pose estimation network based on feature pyramid network structure with attention-based detection modules. This network can detect multiple humans in more than 60 fps with 384 x 384 resolution on GTX 1070 with affordable accuracy. This work shows the potential of this network structure can perform both faster and better compared with state-of-the-art results.