[姿态预测] Flowing ConvNets for Human Pose Estimation in Videos

    xiaoxiao2021-03-26  40

    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: 

    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 

    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

    通过综合前后帧信息来提高鲁棒性。

    结果: 

    转载请注明原文地址: https://ju.6miu.com/read-350041.html

    最新回复(0)