弱监督语义分割

• what: segment image only use image-level label
• challenge: how to build relation between image labels and pixels

参考

1. Features

• 针对输入图片的salience map
  
  • gradient based
• 针对最后一层的class activation map
  
  • 直接feature层面的加权
• excitation backpropagation

2. salience map based

• source:

  • simple images (背景干净只含有单个类别目标)
  • complex images(背景嘈杂并含有多类目标)

• challenge:

  • classification networks are only responsive to small and sparse discriminative regions from object of interest
  • how to get dense and intergral regions

• method:

  • graph cut

Papers

work: Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

pro

1. a bunch gradient based methods, simple and effective
2. a method for gettting salience map of single image
3. a method for weakly segmentation (salience map + graph cut)

con

1. salience based on one output class
   
   • how to combine multiple classes
2. 行文结构较散