CoTraining and its Uses

1
Co-Training and its Uses

• Unsupervised Improvement of Visual Detectors
  using Co-Training (ICCV '03)?
• A. Levin, P. Viola, Y. Freund
• Hebrew Univ. MSR Columbia Univ.
• Online Detection and Classification of Moving
  Objects Using Progressively Improving Detectors
  (CVPR '05)?
• Omar Javed, Saad Ali, Mubarak Shah
• University of Central Florida

2
Co-Training

• Classifiers need labeled data, but
• data is expensive. So,
• Have classifiers teach each other!

3
Co-Training Visualized
4
Levin, Viola, Freund's Work Idea

• Want to detect cars on a road, used to count of
  cars that pass an area
• Build classifiers on training set, find threshold
  above which all samples correctly labeled
• On the unlabeled data, any sample which falls
  above the threshold is considered labeled and
  passed to the other classifier

5
Feature and Classifier Types

• Appearance information is encoded using a
  gray-level histogram
• Camera is stationary, use background subtraction
  to find motion regions. These regions serve as
  the other feature (ie is there motion, and what
  does it look like)
• A LogAdaBoost (Collins) classifier is trained for
  each using 50 positive samples
• 22,000 unlabeled samples are also fed in and
  incorporated using co-training

6
Co-Training for Detection

• After initial train, find ?p and ?n. Label some
  unlabled data, find new ?p and ?n, repeat
• Are many possible negative samples, pick those
  closest to ?n
• For a pos sample, number of windows have high
  score, only pick the local max
• Let it run loose on some data! Performance won't
  be higher than previous methods, improvement is
  in small of labeled data

7
Results (kinda)?

• ROC for Gray-level feature (left) and motion
  feature (right).Green lines are before
  co-training. Black lines are after
  co-training.
• Gray-level and background subtraction detection
  results before and after co-training

8
UC Florida's Work The Idea

• Train an initial detector. But this detector cant
  achieve its maximum performance if its too
  general
• Perform an online update of the classifier.
  Classifier will learn to perform well in its
  specific environment

9
Feature and Classifier Types

• Perform PCA, keep m largest eigenvectors. Do this
  for pedestrian and for vehicle data
• Feature vector is rTSmp, rTSmv where r is the
  sample (in column-vector form, size dx1 where d
  of pixels) and S is the dxm eigenvector matrix
• Construct a Bayes classifier for each coefficient
  of the feature vector

10
How Co-Training is Used

• Classifiers are trained using AdaBoost, where the
  base classifier is the Bayes classifiers
• Find the ?p and ?n, label the samples. Make more
  conservative by enforing that 1/10 of base
  classifiers must confidently label a sample in
  order to add it to the classifier
• Only keep negative samples close to ?n

11
Results

• 50 labeled samples used. Put cameras in 3 areas
  and let the online update run for each

12
Co-Tracking! (is awesome)?

• Start with a few labeled frames
• Build an initial classifier, one for each feature
  type
• In the first unlabeled frame, use your
  classifiers to find the object
• Each classifier gets a vote to the final tracking
  result

13
Feature and Classifier Types

• Color Histogram often very discriminative, but
  performs poorly over illumination changes
• Histogram of Gradients (HoG) often not as
  discriminative, but (mostly) invariant to
  illumination changes
• Features are complimentary. And usually at least
  one will find the object

14
How Co-Training is Used

• After classifiers vote and a final location is
  found this frame is labeled. Add the object
  location as positive sample if confidence is
  high
• Need negative samples, but tradeoff of certainty
  vs. usefulness
• Find BG regions HoG did poorly, pass them as
  negative samples to color classifier, and vice
  versa

15
Results!
16
Thank You!