Dynamic Cascades for Face Detection

Abstract

In this paper, we propose a novel method, called "Dynamic Cascade", for training an efficient face detector on massive data sets. There are three key contributions. The first is a new cascade algorithm called "Dynamic Cascade", which can train cascade classifiers on massive data sets and only requires a small number of training parameters. The second is the introduction of a new kind of weak classifier, called "Bayesian Stump", for training boost classifiers. It produces more stable boost classifiers with fewer features. Moreover, we propose a strategy for using our dynamic cascade algorithm with multiple sets of features to further improve the detection performance without significant increase in the detector's computational cost. Experimental results show that all the new techniques effectively improve the detection performance. Finally, we provide the first large standard data set for face detection, so that future researches on the topic can be compared on the same training and testing set.

A survey on face detection in the wild: Past, present and future

“…Since a reasonable goal make not be known a priori, the algorithm adjusts its cost function depending on the attainability of the goal based on cost prediction. In [68], a dynamic cascade was proposed, which assumes that the false negative rate of the nodes changes exponentially in each stage, following the idea in [106]. The approach is simple and ad hoc, though it appears to work reasonably well.…”

“…To deal with that it was suggested to use fine granularity in the first few layers of the cascade, and coarse granularity in latter layers. Another interesting recent method was proposed in [68], where a new weak classifier called Bayesian stump was introduced. Bayesian stump is also a histogram based weak classifier, however, the split thresholds of the Bayesian stump are derived from iterative split and merge operations instead of being at equal distances and fixed.…”

“…In addition, later stages of the training will be slow due to the increasing number of negative examples, since none of them will be removed during the process. In [68] and [121], the use of importance sampling in order to address the large data set issue, was proposed. The training positive or negative data set were periodically re-sampled to ensure feasible computation.…”

Face Alignment Using Boosting and Evolutionary Search

“…Following the method in [12], we can easily build a k-bins Bayesian Stump. Moreover, we can extend it to a Look Up Table(LUT) weak classifier for RealBoost algorithms by using log-likelihood output to replace the binary output in every interval.…”

“…Xiao et al [12] proposed a method called Bayesian Stump to find P (ω c , x), c ∈ {1, 2} by using histogram to estimate the probability distribution. We divide all features' output value {μ(θ i )} into k sections δ k = (r k−1 , r k ], and the histogram of P (ω c , x) is…”