Strategies for improving face recognition from video

1
Strategies for improving face recognition from
video
Deborah Thomas, Nitesh V. Chawla, Kevin W.
Bowyer, and Patrick J. Flynn Computer Vision
Research Lab, University of Notre Dame
(http//www.nd.edu/cvrl)
Dataset
Goals

• Improve performance of face recognition from
  video
• Exploit multiple frames available in a given clip
• Select a minimal set of frames to represent the
  subject

• Notre Dame dataset
• 105 subjects gallery indoor, probe outdoor
• 3 cameras HD JVC, DV Canon, iSight Webcam
• Honda/UCSD dataset 20 subjects, 1- 4 clips each

Using Mahalanobis cosine distances in PCA space
to determine diversity
Using K-means clustering to group similar images

• Hypothesis The Mahalanobis cosine distance
  between images in PCA space reflects their
  difference
• Process
• Project images from subject into PCA space
• Pick image with largest total Mahalanobis cosine
  distance from all others
• Pick successive images that are farthest away
  from the previously selected image

• Hypothesis Image clusters in PCA space represent
  images that are similar to each other
• Process
• Project images from subject into PCA space
• Use retained dimensions to cluster images
• One image per cluster for N-frame representation

Combining quality measure with difference

• FaceIts faceness measure Confidence that image
  contains a face
• Three approaches
• LAD Use Mahalanobis cosine distance to determine
  distance, use N frames most different from each
  other
• LADHF Project top 35 frames with highest
  faceness into PCA space, use N frames most
  different from each other
• CLS Create K clusters of images and pick image
  with highest faceness from N clusters for
  representation
• Experiments
• Use up to 20 frames
• Compare to choosing images that are equally
  spaced in time

Example images Notre Dame Dataset
Results Canon as probe and gallery
Gallery Outdoor
Probe Indoor
DV Canon
HD JVC
iSight Webcam
Results UCSD/Honda dataset

• Lee et al.
• Use clusters in PCA space to determine pose
• Posterior probabilities Identity of current
  frame conditioned on previous frames
• Rank One recognition rate
• 2003 92.1
• 2005 98.9
• Our approach
• Using 7 frames 98.8

Acknowledgements Biometrics research at the
University of Notre Dame is supported bythe
National Science Foundation under grant CNS
0130839, by theCentral Intelligence Agency, by
the National Geo-Spatial IntelligenceAgency, by
UNISYS Corp., and by the US Department of
Justice.