Pose Machines provide a powerful modular framework for articulated pose estimation. The sequential prediction framework allows for the learning of rich implicit spatial models, but currently relies on manually designed features for representing image and spatial context. In this work, we incorporate a convolutional network architecture into the pose machine framework allowing the learning of representations for both image and spatial context directly from data. The contribution of this paper is a systematic approach to composing convolutional networks with large receptive fields for pose estimation tasks. Our approach ad- dresses the characteristic difficulty of vanishing gradients during training by providing a natural learning objective function that enforces intermediate supervision, thereby replenishing backpropagated gradients and conditioning the learning procedure. We demonstrate state-of-the-art performance and outperform competing methods on standard benchmarks.

本研究提出了一种通过将卷积网络与姿态机器学习框架相结合的系统设计来学习图像特征和图像相关的空间模型，以实现姿态估计的任务，并解决了训练过程中梯度消失的问题，其通过提供自然的学习目标函数来约束学习过程，最终在MPII， LSP和FLIC等标准基准测试中实现了最先进的性能表现并超越了竞争方法。

卷积姿势机