Auto-Context and Its Application to High-level Vision Tasks
Zhuowen Tu CVPR 2008Presented by Vladimir Reilly
Problems Tackled in Paper Horse Segmentation
Label Every pixel in image as horse or background
Problems Tackled in Paper Image labeling
More complex segmentation
Problems Tackled in Paper Human body Segmentation
Label Body Parts
Solution Context ADABOOST
Cool Idea Contextual information is integrated directly into
ADABOOST Context not limited by spatial proximity Fast General
Context
Appearance Context
Label Context
?
Tree?
Grass?
Sky?
Human?
Grass
Previous Work CRFs
Previous Work Spatial Boost
In addition to appearance InformationLook at labels of neighbor pixels
Derive weak Spatial Learner
The Algorithm Iteration 1
Train Image Label Map
Extract 21x21patch
Generate Weak Appearance
Learners8000 possible features
Train Strong Classifier
The Algorithm Iteration > 1
Train Image Label Map
Segment Images
Probability Map
Extract 21x21patch
Generate Weak Appearance
Learners8000 possible features
Generate Weak Context Learners
Shoot RaysSample Along RaysCompute Statistics
4000 possible features
Probability out of adaboost
PBT
Results
Results
Google Images
Interesting Observations Starting with second classifier
90% of selected learners are context learners Label Context improves results Appearance Context worsens results
Probability Map
Train Image
Results
Results
Results