Session 7: Face Detection (cont.) John Magee 8 February 2017 Slides courtesy of Diane H. Theriault
Question of the Day: How can we find faces in images?
Face Detection Compute features in the image Apply a classifier Viola & Jones. Rapid Object Detection using a Boosted Cascade of Simple Features
What do Faces Look Like? Chosen features are responses of the image to box filters at specific locations in the image
Using Features for Classification Gateway between the signal processing world and the machine learning world For any candidate image, we will compute the response of the image to some different filters at some different locations These will be our input to our machine learning algorithm
Using Features for Classification In Machine Learning, a classifier is a thing that can make a decision about whether a piece of data belongs to a particular class (e.g. is this image a face or not ) There are different types of classifiers, and one way to find their parameters by looking at some training data (i.e. supervised learning ) A weak learner is a classifier for which you can find the best parameters you can, but it still does a mediocre job
Using Features for Classification Weak Learner Example: Apply a threshold to a feature (the response of the image to a filter) How to find a threshold?
Using Features for Classification How to find a threshold? Start with labeled training data 9,832 face images (positive examples) 10,000 non-face images (sub-windows cut from other pictures containing no faces) (negative examples) Compute some measure of how good a particular threshold his (e.g. accuracy ), then find the threshold that gives the best result
Using Features for Classification For a particular threshold on a particular feature, compute: True positives (faces that are identified as faces) True negatives (non-face patches that are identified as non-faces) False positives (non-faces identified as faces) False negatives (faces identified as non-faces) Accuracy: % correctly classified Classification Error: 1 - Accuracy For each feature, choose a threshold that maximizes the accuracy Classifier Result positive negative Known Classification positive True Positive False Negative negative False Positive True Negative Confusion Matrix
Using Features for Classification How do you know which feature to use? Try them all and pick the one that gives the best result Then, choose the next one that does the next best job, emphasizing the misclassified images Each threshold on a single feature gives mediocre results, but if you combine them in a clever way, you can get good results (That s the extremely short version of boosting )
Training: Classification with Adaboost Given a pool of weak learners and some data, Create a boosted classifier by An awesome Machine Learning Algorithm! choosing a good combination of K weak learners and associated weights In our case Train a Weak Learner = = Choose a feature to use and which threshold to apply
Classification with Adaboost Training: Initialization: Assign data weights uniformly to each data point For 1:K Train all of the weak learners Compute the weighted classification error using weights assigned to each data point Choose the weak learner with the lowest weighted error Compute a classifier weight associated with the weak learner, based on the classification error Adjust the weights for the data points to emphasize misclassified points (Specifics of how to compute weights in paper)
Classification with Adaboost Classification: Use the boosted classifier (the weak learners and associated weights we found during training) to label faces Evaluate each weak learner we chose on the new data point by computing the response of the image to the filter and applying the threshold to obtain a binary result Make a final decision by computing a weighted sum of the classification results from the weak learners
Classification Cascade Tradeoff between more complex classifier that uses more features (computational cost) and accuracy What is an acceptable error rate? What is an acceptable computational cost? Can we have our cake and eat it too?
Classification Cascade Solution: Use a cascade of increasingly complex classifiers Create less complex classifiers with fewer weak learners that achieve high detection rates (maybe with extra false positives) Evaluate more complex, more picky, classifiers only after the image passes the early classifiers Train later classifiers in the cascade using only images that pass earlier classifiers
To Detect Faces Divide large images into overlapping sub-windows Apply classifier cascade to each sub-window Apply to sub-windows of different sizes by scaling the features (using larger box filters)
Discussion Questions: What is the relationship between an image feature and the response of an image to a box filter applied at a particular location? If you were given a set of labeled images and a filter response for some particular filter, how would you choose a threshold to use? How would you adjust your procedure for finding the best possible threshold if you wanted to find the best threshold that recognized at least 99% of faces, even if it let through some nonfaces (false positives)? Given an image, what are the steps for labeling it as face or nonface? What is a classifier cascade and why would you want to use one?