A mixed generative-discriminative framework for pedestrian classification

Abstract

This paper presents a novel approach to pedestrian classification which involves utilizing the synthesized virtual samples of a learned generative model to enhance the classification performance of a discriminative model. Our generative model captures prior knowledge about the pedestrian class in terms of a number of probabilistic shape and texture models, each attuned to a particular pedestrian pose. Active learning provides the link between the generative and discriminative model, in the sense that the former is selectively sampled such that the training process is guided towards the most informative samples of the latter.In large-scale experiments on real-world datasets of tens of thousands of samples, we demonstrate a significant improvement in classification performance of the combined generative-discriminative approach over the discriminative-only approach (the latter exemplified by a neural network with local receptive fields and a support vector machine using Haar wavelet features).

Monocular Pedestrian Detection: Survey and Experiments

“…To recover physically plausible regions in the linear model space, conditional density models have been proposed [9], [14]. Further, nonlinear extensions have been introduced at the cost of requiring a larger number of training shapes to cope with the higher model complexity [9], [14], [25], [26], [50]. Rather than modeling the nonlinearity explicitly, most approaches break up the nonlinear shape space into piecewise linear patches.…”

“…Forcing topologically diverse shapes (e.g., pedestrian with feet apart and with feet closed) into a single linear model may result in many intermediate model instantiations that are physically implausible. To recover physically plausible regions in the linear model space, conditional density models have been proposed [9], [14]. Further, nonlinear extensions have been introduced at the cost of requiring a larger number of training shapes to cope with the higher model complexity [9], [14], [25], [26], [50].…”

“…Continuous shape models involve a compact parametric representation of the class-conditional density, learned from a set of training shapes, given the existence of an appropriate manual [9], [25], [26] or automatic [4], [5], [14], [34], [50] shape registration method. Linear shape space representations which model the class-conditional density as a single Gaussian have been employed by Baumberg [4] and Bergtholdt et al [9].…”

A system for large-scale automatic traffic sign recognition and mapping

“…It was successfully used in many applications, such as: human pose recognition (Shotton et al, 2011), object 3D structure inferring (Grauman et al, 2003), shape models learning (Stark et al, 2010), pedestrian detection (Marin et al, 2010) (Pishchulin et al, 2011) (Enzweiler et al, 2008), viewpoint-independent object detection (Liebelt et al, 2008), text recognition (Wang et al, 2011) and keypoints recognition (Ozuysal et al, 2007).…”

In Good Shape: Robust People Detection based on Appearance and Shape

“…Similarly, Marin et al [22] use a game engine to generate training samples from multiple viewpoints. Broggi et al [5] aimed to synthetically model pedestrians' appearance for detection in infrared images, while [10] varied 2D pedestrian appearance by employing a generative model for shape and texture of pedestrians. Recently, Pishchulin et al [23] use a 3D computer graphics model to re-render images in order to obtain a larger number of synthetic training samples with realistic appear-ance.…”