1. Limitations of Cotemporary Classification Algorithms Major limitations of classification algorithms like Adaboost, SVMs, or Naïve Bayes include, Requirement of a large amount of labeled training data. Fixed parameters after the end of training phase, i.e., these classifiers can not attune themselves to particular detection scenarios after deployment. Synopsis A boosted classification framework, in which, Separate views (features) of the data used for online collection of training examples through co-training. Combined view (all features) used to make classification decisions. Background modeling used to prune away stationary regions to speed up the classification process. Global feature representations used for robustness. Online Detection and Classification of Moving Objects Using Progressively Improving Detectors Omar Javed, Saad Ali, Mubarak Shah Computer Vision Lab University of Central Florida Initial Training The Bayes Classifier is used as the base (weak) classifier for boosting. Each feature vector component v q,where q ranges from 1,.., m1+m2 (for two object classes + background class), is used to learn the pdf for each class. The classification decision by the qth base classifier h q is taken as c i, Adaboost.M1 (Freund and Schapire, 96) is used to learn the strong classifier, from the initial training data and the base classifiers. Results Initial Training: 50 examples of each class All examples scaled to 30x30 vector Validation Set :20 images per class Testing on three sequences (a) Change in performance with increase in time for sequence 1,2 and 3. The performance was measure over two minute intervals. Over 150 to 200 detection decisions were usually made in this time period. Properties of the Proposed Algorithm Requirement of minimal training data Automated online selection of training examples after deployment, for continuous improvement.. Near real time performance (4-5 frames/sec) The Online Co-training Framework During the test phase, select example for training, if more than 10% of the classifiers confidently predict the label of an example. Example’s margin is less than the computed thresholds. Once an example has been labeled by the co-training mechanism, an online boosting algorithm by [Oza and Russel,02] is used to update the base classifiers and the boosting coefficients. Feature Extraction Features for classification are derived from Principal Component Analysis of the appearance templates of the training examples. For each object class c i (excluding background) an appearance subspace, represented by d x m i projection matrix S i, is constructed. m chosen such that eigenvectors represent 99% of respective subspace. Appearance features for base learners are obtained by projecting a training example ‘r’ into appearance subspace of each object class. for two object classes the feature vector v of an example will be, Probability Vehicle Person Clutter Histograms of a feature coefficient from the appearance subspace Row 1: Top 3 eigenvectors for person appearance subspace. Row 2: Vehicle appearance subspace If Where β {i=1…N} are the boosting parameters, C is the set of classes The Online Co-training Framework : Key Observations Boosting mechanism selects the least correlated base classifiers. Ideal for co-training! Examples confidently labeled by one classifier are used to train the other. Only the observations lying close to the decision boundary of the boosted classifier are useful for improving classification performance. Use examples with small margins for online training. The Online Co-training Framework: Implementation Steps Determine confidence thresholds for each base classifier, using a validation data set. For class c i and jth base classifier h j set the confidence threshold, T j,ci base ( highest probability achieved by a negative example). Compute Margin thresholds T ci base are from the validation data set. Foreground Models FeatureExtraction ROIs Background Background Models Updated parameters Color Classifier Edge Classifier Base Classifiers Boosted Classifier Classification Output Updated Boosted Parameters Co-Training Decision (if confident prediction by one set) Updated weak learners Information flow for the real-time object classification system Performance vs. the number of co-trained examples for the three sequences. Relatively few examples are required for improving the detection rates since these examples are from the same scene in which the classifier is being evaluated.