Flowing ConvNets for Human Pose Estimation in Videos

ICCV 2015 code available
http://www.robots.ox.ac.uk/~vgg/software/cnn_heatmap/

本文主要用CNN网络来进行人体姿态估计,加入了temporal 信息以提高精度。
网络框架如下:
Flowing ConvNets for Human Pose Estimation in Videos_第1张图片

本文对于关节位置的估计提出了一个 heatmap概念,而不是一个坐标的回归。这样做可以提高关节定位的鲁棒性。

Spatial fusion layers 这主要是用来提取关节之间内在联系的。
learn dependencies between the human body parts locations represented by
these activations
Flowing ConvNets for Human Pose Estimation in Videos_第2张图片

Optical flow for pose estimation: 使用光流法来增强 heatmaps,具体通过以下三个步骤来实现:
1) the confidences from nearby frames are aligned to the current frame using
dense optical flow
2) these confidences are then pooled into a composite confidence map using
an additional convolutional layer
3) the final upper body pose estimate for a frame is then simply the positions
of maximum confidence from the composite map

Flowing ConvNets for Human Pose Estimation in Videos_第3张图片
通过综合前后帧信息来提高鲁棒性。

结果:
Flowing ConvNets for Human Pose Estimation in Videos_第4张图片

你可能感兴趣的:(行人检测,行人属性,ZJ)