Hausdorff Forum November 3, 2017 - 14:15

Location: Lipschitz-Saal, Endenicher Allee 60

14:15 Jürgen Gall (Universität Bonn): "Analyzing Human Behavior in Video Sequences"

Analyzing the behavior of humans in continuous video recordings requires to solve several tasks. This includes the estimation and tracking of poses of multiple persons or the temporal detection of activities. In this talk, I will describe some of our recent works for this two tasks. The first task requires to solve jointly the problem of person association over time and the pose estimation for each person. This problem can be formulated as a graph partitioning problem where a spatio-temporal graph is constructed from detected body joints in a video. For the second task, temporal models like recurrent neural networks are usually trained on videos that are annotated at a frame-level. Acquiring such annotations, however, is very time consuming and strong temporal models require large amounts of annotated training data. Weaker forms of supervision like transcripts are therefore investigated to learn temporal models.

15:15 - 15:45 Tea

15:45 William A. P. Smith (University of York, UK): "Model-based analysis of faces"

The quest to understand and model "face space" dates back to the 1980s, though the variability and uniqueness of faces has fascinated scholars, artists and scientists since antiquity. To learn a face space from a sample of face data requires the factors that are intrinsic to the face to be disentangled from extrinsic factors related to the imaging environment. A model-based approach to analysis of face images uses explicit models of the geometric and photometric image formation processes in order to explain an image in terms of factors such as shape, lighting and skin reflectance properties. This is in contrast to learning-based approaches where a black box (usually a convolutional neural network) is trained to directly classify or regress some property of interest from an image. In this talk, I will present a variety of work on model-based analysis of faces, describe current work to try to integrate explicit models into black box learning, present some applications including evaluation of craniofacial surgical outcomes and discuss a collaboration with psychologists to try to uncover the models and representations used in human perception of faces.