Cascaded Models for Articulated Pose Estimation Ben Sapp,
Alexander Toshev and Ben Taskar University of Pennsylvania
Slide 2
Human Pose Estimation Goal: Image -> Stick Figure 2D
locations of anatomical parts from a single image efficient
inference input output
Slide 3
Human Pose Estimation: Its Hard lighting variation background
clutter pose variation intrinsic scale variations
foreshortening
Slide 4
Articulated Pose and Pictorial Structures A popular choice for
(articulated) parts-based models A non-exhaustive timeline Fischler
& Elschlager The representation and matching of pictorial
structures Felzenszwalb & Huttenlocher PS for Object
Recognition Ramanan Learning to Parse Images of Articulated Objects
Felzenszwalb et al. A Discriminatively Trained, Multiscale,
Deformable Part Model. Ferrari et al. Progressive Search Space
Reduction 1972200520062008 2009 2010 Eichner & Ferrari Better
Appearance Models for Pictorial Structures Andriluka et al.
Pictorial Structures Re-visited Fergus et al. ICCV Short Course
Sapp et al. Adaptive Pose Priors for PS
Slide 5
Background: How PS works x y head part detectors unary score:
detection maps pairwise score: geometric prior llarm rlarm ruarm
luarm torso : location for part i max-product inference sum-product
inference prediction
Slide 6
Background: The Complexity of PS state space for part i typical
state space size: n > 150,000 states x y head llarm rlarm ruarm
luarm torso Standard inference in a tree graphical model is Typical
# of valid combinations for two neighboring parts: (80 x 80 x 24)
(80/5 x 80/5 x 24) 1 billion state-pairs! x = pairwise
computation:
Slide 7
Background: The Complexity of PS If, efficient inference tricks
can be used: [Felzenszwalb & Huttenlocher, 2005 ] Max-prod w/
unimodal cost: Distance transform for Sum-prod w/ linear filter
cost: Convolution for Q: Are we losing too much in expressivity for
this gain in efficiency? + score for part-state pair: + unary i
unary j pairwise i,j State-of-the-art:
Slide 8
Goal: Integrating richer pairwise terms (, ) e.g., distance in
color distribution: incorporate image evidence