Image-level to Pixel-wise Labeling: From Theory to Practice[IJCAI18]

Image-level to Pixel-wise Labeling: From Theory to Practice

• incorporate image label in segmentation network
• use generative network to recover original image and constrain

• analyze effects of image labeling on pixel-wise segmentation: accurate and noise

• maybe we can use first two.

Methodology

fix image-level network, cotrain segmenation and generative network

image-level

use a pretrained network

segmenation

FCN

gernerative

3 conv layers

trainning procedure

each iteration has three steps: smoothness in multi-network optimization
1. the segmentation network is trained individually
2. generative network is trained to reconstruct the input image with the parameters of
the segmentation network fixed.
3. the segmentation and generative networks are jointly trained