On the design of robust classifiers for computer vision

Abstract

Abstract

TaylorBoost: First and second-order boosting algorithms with explicit margin control

“…A classifier is trained to distinguish the target object from the background in each video frame, and used in the next frame for tracking [2]. Various methods have been proposed to learn the classifier, including AdaBoost [2], discriminant saliency [8], and a combination of discriminant saliency and TangentBoost [11]. The latter achieved the best results in the literature.…”

“…Boosting is a reliable tool for this design. Since the introduction of AdaBoost in [5], a number of algorithms have appeared in the literature, including LogitBoost [6], GentleBoost [6], GradientBoost [12], or TangentBoost [11]. They all minimize a risk that upper bounds the classification error, and converge asymptotically to the Bayes decision rule.…”

A Correction on Tangentboost Algorithm

“…Thus, outliers and contaminated part in training data can spoil the classi…ers in boosting methods. Great advances have been achieved to make more robust boosting algorithms in the last decade [4,5,6,7,8].…”

Structured Learning for Multiple Object Tracking

“…It only uses the reliable samples from the first frame as labeled, and considers the one from other frames as unlabeled. [16,18] incorporate more robust loos functions in boosting. Multiple Instance Learning (MIL) is used by [3,23] to automatically elicit the best positive sample during training.…”

“…The key to overcome the difficulty is to adaptively update the appearance model during tracking, c.f., [3,9,10,13,16,18,19,20,21,23], which will be referred as adaptive tracking-by-detection methods in this paper. These methods use previous tracking results to generate a new training set for object appearance, and update the current model to predict the object location in subsequent frames.…”