YujunXie / Papers-of-Crowd-Counting

0 stars 1 forks source link

CPN #2

Open YujunXie opened 4 years ago

YujunXie commented 4 years ago

CPN

《Cascaded Pyramid Network for Multi-Person Pose Estimation》

简介

方法

image-20191130100144132

The localization of keypoints like hip usually requires more context information and processing rather than the nearby appearance feature.

top-down方法,即先检测人体后预测关键点。resnet+globalnet+refinenet

GlobalNet:由特征金字塔FPN预测简单的关键点,得到丰富的语义特征。浅层特征空间信息丰富,深层特征语义信息丰富,两者信息互补(可视化VGG-16的特征图可得出相同结论)。

RefineNet:通过精细化的级联网络回传在线困难关键点的损失,由此预测难的关键点。

实验

思考:thinking:

:dart: take-home-message

:sparkles: 对其他任务的启发

参考文献:books:

OHEM

《Feature pyramid networks for object detection》

《Learning feature pyramids for human pose estimation》

《Training region-based object detectors with online hard example mining》

《Mask R- CNN》

《A coarse-fine network for keypoint localization》

YujunXie commented 4 years ago

《Multi-Person Pose Estimation with Enhanced Channel-wise and Spatial Information》

简介

方法

image-20191130105434550

The channel information with different characteris- tics among different levels can complement and reinforce with each other.

  1. 不同level的特征图上采样后concat, 经过channel shuffle通过分组卷积并下采样后回到之前的大小。
  2. 将混洗前后的特征图concat,得到金字塔特征图送入RefineNet。

其中channel shuffle操作参考shuffleNet:对于一个卷积层分为g组,

1.有g×n 个输出通道

2.reshape为(g,n)

3.再transpose为(n,g)

4.flatten,再分回g组作为下一层的输入


  1. 空间注意力机制与通道注意力机制(SENet)
image-20191201104009205

实验

思考:thinking:

:question:结果对比分析各个指标都有提升,与CPN相比体现了特征信息提萃的优点。但是没有模型参数的对比。

:dart: take-home-message

:sparkles:

参考文献:books:

轻量级网络--ShuffleNet论文解读

《Cascaded pyramid network for multi-person pose estimation》

《Squeeze-and-excitation networks》

《Efficient multi-scale training》

《Simple baselines for human pose estimation and tracking》

《SCA-CNN: Spatial and Channel-wise Attention in Convolutional Networks for Image Captioning》