论文原文:https://arxiv.org/pdf/1603.06937.pdf 作者信息:Alejandro Newell, Kaiyu Yang, and Jia Deng 项目地址:https://github.com/princeton-vl/pose-hg-train 博客参考:https://blog.csdn.net/JerryZhang__/article/details/98308729

image.png
网络特点:堆叠的沙漏结构(多分辨率) + 中继监督
image.png
多分辨率的优势:

One consideration is that for many failure cases a refinement of position within a local window would not offer much improvement since error cases often consist of either occluded or misattributed limbs. For both situations, any further evaluation at a local scale will not improve the prediction.

image.png
向上采样方式:插值算法,反卷积算法等。本文采用最邻近插值,然后结合底层的 feature maps 实现。
image.png

此处的中继监督和CPM网络的中继监督是类似的,都是产生 predcitions 之后再次连接回主网络中。仔细想一下的话,两者的结构也是类似的,都是分级结构,但是具体实现的方式不同。

结果

image.png
image.png
如何验证:对正常图像和翻转的图像进行估计之后取平均值(AlphaPose也是如此ECCV_2016_Stacked Hourglass Network - 图7常用ECCV_2016_Stacked Hourglass Network - 图8能够提升预测的准确率)
堆叠结构(级联结构)的有效性?
image.png