Region Mutual Information Loss for Semantic Segmentation

Explore a 21-minute conference talk from the University of Central Florida on Region Mutual Information Loss for Semantic Segmentation. Delve into key concepts like convolutional-ization, encoder-decoder architecture, à trous convolutions, and CRF post-processing. Learn about multi-scale features, à trous spatial pyramid pooling, and DeepLabv3+ failure cases. Understand the motivation behind Region Mutual Information Loss, examining mutual information from entropy and probability perspectives. Discover the overall loss function, experimental setup, datasets used, and results on PASCAL VOC 2012, including ablation studies. Access accompanying slides for visual aids and additional information.

Syllabus

tl;dr / one line summary
Sections
Semantic Segmentation
Key Idea: "convolutional-ization"
Key Idea: Encoder-decoder architecture
Key Idea #1: Convolutionalization with "à trous convolutions"
Key Idea #2: CRF Post-Processing
Key Idea: Multi-Scale Features
Key Idea: Deeper with à trous convolutions
Key Idea: À trous Spatial Pyramid Pooling
DeepLabv3+: Failure Cases
Motivation for Region Mutual Information Loss
Mutual Information: Entropy View
Mutual Information: Probability View
Overall Loss Function
Experimental Setup
Datasets
Results: PASCAL VOC 2012
Ablation Studies: PASCAL VOC 2012