Head Pose Estimation in Single

1
Head Pose Estimation in Single- Multi-view
Environments

• Results on the CLEAR07 Benchmarks
• Michael Voit, Kai Nickel, Rainer Stiefelhagen
• Universität Karlsruhe

2
System Overview

• System uses
• Single-view Head Pose Estimation (Neural Network)
• Fusion of all single hypotheses
• (Bayesian Filtering ? Tracking)

3
Single-view Head Pose Estimation

• NN Input
• Cropped Head Region
• Resized
• Grayscaled
• Enhanced Contrast
• Edge Image
• NN Output
• Discrete distribution of head rotations

4
Multi-view Fusion and Head Pose Tracking
Compute current, merged observation (average
single-views likelihoods per angle)
Diffuse Distribution (Convolution with Gaussian
Kernel)
Compute new posterior Distribution
5
Results AMI Corpus (single-view)
6
Results CHIL Corpus (multi-view)
7
Interesting to know...

• Fusing views helped to reduce error by 90
• Compared to choosing the most likely angle over
  all views (on a per frame basis)
• Transition modeling included reduced error by 2

8
Thank You

• Any Questions?

9
Multi-view Fusion and Head Pose Tracking

• Final, joint hypothesis is derived after applying
  a Bayesian Filter (equals Particle Filter w/o
  resampling)
• Fixed set of states (360 states for 360 pan
  angles, 180 for 180 tilt angles)
• State likelihood by averaging corresponding
  single-views likelihoods
• Transitioning is modeled convoluting previous
  posterior distribution with gaussian kernel
  (StdDev54)

10
Multi-view Fusion and Head Pose Tracking

• Final, joint hypothesis is derived after applying
  a Bayesian Filter (equals Particle Filter w/o
  resampling)
• Fixed set of states (360 states for 360 pan
  angles, 180 for 180 tilt angles)
• Single-views likelihoods are averaged to obtain
  merged observation

Likelihood
...
359
1
2
3
358
0
Pan angle