Title:
2D Human Pose Estimation and Search in (almost) Unconstrained Video

Abstract:
The first part of this talk tackles fully automatic 2D human pose estimation
in (almost) unconstrained TV shows and feature films. Direct pose estimation
on this uncontrolled material is often too difficult, especially when
knowing nothing about the location, scale, and appearance of the person. We
extend traditional pictorial structure pose estimators with mechanisms to
progressively reduces the search space for body parts. These greatly
facilitate the task of pictorial structure fitting and allows to deal with
substantially more challenging images and videos.
The second part of the talk presents a new retrieval application made
possible by the better pose estimator: retrieving shots containing a
particular pose from a video database. The user can specify a query pose
through a single frame, or through a small set examples. The method
describes the spatial configuration of body parts returned by the pictorial
structure with features which are person, clothing, background and lighting
independent. This allows to generalize well even when starting from a single
query frame.
We compare our approach quantitatively to a baseline using HOG descriptors,
and show results for various poses on several episodes of `Buffy the Vampire
Slayer' and Hollywood movies.