- InfoSeg Unsupervised Semantic Image Segmentation with Mutual Information Maximization
- Motivation
- Introduce a novel unsupervised semantic segmentation based on mutual information maximization between local and global high-level features
- Background
- Color segmentation fails to learn high-level representation
- Representation Learning captures global image-level representation
- InfoSeg learns semantic high-level representation
- Method
- Global image-level features H from the output projection of block D
- Local patch-wise features L from the output projection of block C
- Class-probability volume V
- Define feature assignments S for Mutual Information Maximization
- Mutual Information Maximization Loss:
- RESULTS
- Ablation Study
InfoSeg Unsupervised Semantic Image Segmentation with Mutual Information Maximization
InfoSeg Unsupervised Semantic Image Segmentation with Mutual Information Maximization
InfoSeg: Unsupervised Semantic Image Segmentation with Mutual Information Maximization
Motivation
Introduce a novel unsupervised semantic segmentation based on mutual information maximization between local and global high-level features
Background
Color segmentation fails to learn high-level representation
Representation Learning captures global image-level representation
InfoSeg learns semantic high-level representation
Method