Checking different versions of adaboost implementations
| Code, data and README: (ZIP). |
ROC comparison
Based on a series of five-fold cross validation experiments on 131 positive images and 600+ negative images, as a comparison, the testing AUC (area under ROC curve) is plotted against the number of training examples, for three classifiers obtained by: (1) adaboost that trains the threshold of MAX1 feature in each iteration on reweighted training sample. (2) adaboost that pre-trains the threshold of each MAX1 feature on initial weights. (3) an older matlab implementation of (2).. In total 5 x 3 x 5 cross validation runs are performed for all training methods and training sample sizes. Images are of size 85 * 127 and are roughly aligned.
|
|
| eps |
Each curve includes five points denoting the AUC (area under ROC curve) on testing data using 5, 10, 20, 40 and 80 positive training examples. Training negatives are kept to 160 images for all runs, and the rest of them are used in testing. 90% condifence bounds are also plotted along with the three curves. All the three strong classifiers contain 80 selected weak classifiers.
Templates learned:
In the figure below we show the selected features by different versions of adaboost learner, on a 40-positive-example training set. For clarity, only the first 30 symbolic sketches of Gabor features are shown.
| Adaboost (full) | Adaboost (limited) | Adaboost (old matlab version) | ||
|
|
|
To generate all the above figures, please run "compareAdaboost.m" under the root folder.