Discriminative models for multi

Download Report

Transcript Discriminative models for multi

Discriminative models for multiclass object layout
Chaitanya Desai, Deva Ramanan, and Charless Fowlkes
Won the David Marr Prize at ICCV 2009
Presented by Esa Rahtu
February 22, 2010
Material taken from authors and Matthew Blaschko.
Multi-object localization as
structured prediction
Learn a mapping
Overview
• Learn about the appearance of objects and
the interactions between objects.
• ”End-to-end” learning framework
– Features
– Optimization problem
• Similar to the approach of Blacshko and
Lambert (ECCV 2008)
A “taxonomy” of interactions
• Previous works use these interactions heuristically.
• Proposed method learns the relative importance of
these cues for the classes we have.
Spatial Co-occurrence Features
• Coarsely discretize the space of object
locations
• Build a feature vector based on the spatial
relationship between objects (dij – 7 dims) .
Regression Problem
• Input space, space of images
• Output space, Y: discretized set of bounding boxes
with an assignment of a class (or none):
Optimization Problem
• Framed as a structured output learning task they use structured output SVM
S ( X i , Yi )  S ( X i , Y )  (Yi , Y )  i
Inference problem
• greedy optimization - heuristic, no
convergence guarantees (!)
Results
Average precision on
VOC 2007
Results
Results
Comments
• Overall idea is very good. They solve actually
much harder problem than other detectors.
• Inference phase is the weakest part (authors
claim ”end to end” framework, but plenty of
ad-hoc solutions are used. (Theoretical
arguments for greedy optimization are not
necessarily good).
• Training time might be long?