Multiview Head Pose Estimation

1
Multi-view Head Pose Estimation

• Evaluation on the CHIL Data Corpus
• Michael Voit
• Universität Karlsruhe

2
Data Overview
3
Data Overview

• Data Corpus consists of
• Trainingset (10 persons)
• Testset (5 persons)
• Each recording per person
• is 3 min. long (_at_15 fps)
• contains 4 camera views
• provides manual annotated head bounding boxes
  (tightest box around skull)
• provides groundtruth head pose (captured with
  Flock of Birds Motion Sensor placed on head,
  30Hz)

4
Metrics

• Mean Error
• Pan
• Tilt
• Roll
• Mean Angular Error (between groundtruth
  estimated vector)
• Cos-Valued
• (Real Angle )

5
Results Mean Error
6
Results Mean Angular Error (Cos)
7
Summary Conclusion

• Using Motion Sensor strongly increased systems
  accuracies compared to CLEAR06 (40 recognition
  rate, pan annotated into 45 wide classes)
• ? Manual annotations were not precise enough or
  annotations were too ambiguous
• Remembering the annotation process for CLEAR06
• Humans even have big (!) problems determining pan
  rotation even in 45 steps ? (Where does South
  start end South-West end?)
• Task solved, more challenges to be included
• Vary Lighting
• Include Occlusions
• Leave out head bounding box annotations (no more
  alignment provided)
• Head Detection AND Tracking to be made
  automatically?
• Deriving Visual Focus of Attention from Head Pose
  Estimates (good targets to track?)