abstract.tex

Human pose estimation from 2D images is one of the most challenging and 
computationally demanding problems in computer vision. Standard models such as 
Pictorial Structures consider interactions between kinematically connected 
joints or limbs, leading to inference cost that is quadratic in the number of 
pixels. As a result, researchers and practitioners have restricted themselves 
to simple models which only measure the quality of limb-pair possibilities by 
their 2D geometric plausibility.

In this talk, we propose novel methods which allow for efficient inference in 
richer models with data-dependent interactions. First, we introduce structured 
prediction cascades, a structured analog of binary cascaded classifiers, which 
learn to focus computational effort where it is needed, filtering out many 
states cheaply while ensuring the correct output is unfiltered. Second, we 
propose a way to decompose models of human pose with cyclic dependencies into a 
collection of tree models, and provide novel methods to impose model agreement.  
Finally, we propose a local linear approach that learns bases centered around 
modes in the training data, giving us image-dependent local models which are 
simple and accurate. 

These techniques allow for sparse and efficient inference on the order of 
minutes per image or video clip. As a result, we can afford to model pairwise 
interaction potentials much more richly with data-dependent features such as 
contour continuity, segmentation alignment, color consistency, optical flow and 
multiple modes. We show empirically that these richer models are worthwhile, 
obtaining significantly more accurate pose estimation on popular datasets.