On the importance of accurate weak classifier learning for boosted weak classifiers

Abstract

Recent work [1], has shown that improving model learning for weak classifiers can yield significant gains in the overall accuracy of a boosted classifier. However, most published classifier boosting research relies only on rudimentary learning techniques for weak classifiers. So while it is known that improving the model learning can greatly improve the accuracy of the resulting strong classifier, it remains to be shown how much can yet be gained by further improving the model learning at the weak classifier level. This paper derives a very accurate model learning method for weak classifiers based on the popular Haar-like features and presents an investigation of its usefulness compared to the standard and recent approaches. The accuracy of the new method is shown by demonstrating the new models ability to predict ROC performance on validation data. A discussion of the problems in learning accurate weak hypotheses is given, along with example solutions. It is also shown that a previous simpler method can be further improved. Lastly, we show that improving model accuracy does not continue to yield improved overall classification beyond a certain point. At this point the learning technique, in this case RealBoost, is unable to make gains from the improved model data.The method has been tested on pedestrian detection tasks using classifiers boosted using the RealBoost boosting algorithm. A subset of our most interesting results is shown to demonstrate the value of method.

Smoothing LUT classifiers for robust face detection

“…It is not impossible because LUT classifiers with minefield partitions are prone to be selected as optimal classifiers on account of their possible small Z values according to (4). Similar observations are also presented in [22,23].…”

“…Meanwhile, the strong classifiers based on smoothed LUT classifiers have a little worse convergent ability than those based on unsmooth ones. Some smoothing schemes for LUT classifiers have been proposed by the researchers, such as smoothing response binning (SRB for short) [18] method and least sigmoid regression [23] (LSR for short) method. The former aims at weight sum smoothing method with adaptive smoothing region and almost all-one smoothing factor, which needs to be extended to more general form for both the smoothing factor and the smoothing width.…”

Boosting a heterogeneous pool of fast HOG features for pedestrian and sign detection

“…These are both fast and discriminative [17] [18]. However, the benefit of improved modelling while keeping the fast lookup table approach has been shown in [19] and [14]. For Haar features this modelling approach yielded a 75% average error reduction [19].…”

“…However, the benefit of improved modelling while keeping the fast lookup table approach has been shown in [19] and [14]. For Haar features this modelling approach yielded a 75% average error reduction [19]. Therefore we apply a similar Smoothed Response Binning Method to our scalar LiteHOG and LiteHOG+ feature responses.…”

“…where x contains the N selected orientations from the histogram image and g() is the RealBoost weak learner classification response using the Smoothed Response Binning approach found in [19].…”