CAMEO: Face Recognition Year 1 Progress and Year 2 Goals

1
CAMEO Face RecognitionYear 1 Progress and Year
2 Goals

• Fernando de la Torre, Carlos Vallespi, Takeo
  Kanade

2
Face Recognition from video.

• How to learn a facial model from the
• data coming from the face detector?

3
Face Recognition from video.

• Challenges
• How to learn INVARIANTLY to spatial
  transformations?
• Simultaneous registration and Subspace
  computation.
• 2) How to select the most discriminative
  features?
• 3) How to deal with missing data?

4
Face Recognition from video.

• Register w.r.t a Subspace
• Selecting the most discriminative samples.

5
Face Recognition from video.
- How to exploit temporal redundancy in the
recognition process?
B
A
Distance between Sets A and B.
Singular vectors of A
6
Face Recognition from video.

• 95 of recognition rate (11 Subjects and 30
  images per subject).

7
Plans year 2.

• Why is hard to perform face recognition from
• Mosaic images?
• Small images.
• Noisy images.
• Misalignments.
• But
• Temporal redundancy.
• Recognizing several people (exclusive principle).
• Superesolution.

8
Learning person-specific models.

• Unsupervised learning from video sequences
• Facial appearance models.
• Behaviour models (e.g. gestures).
• Learning person-specific models can be useful to
  identify people, to predict actions?

9
Meeting visualization/summarization

• Input
• Set of several videos, with detected and
  recognized faces.
• Set of indicators if the person is talking, up,
  down, etc
• Output
• Low dimensional visualization of the meeting
  activity and interaction between people.
• Learning interaction models between people.