Zero-variance minibatch Monte Carlo for pixel-wise visual tracking

Abstract

In this study, the authors present a novel visual tracking method using the pixel-wise posterior estimation and minibatch Monte Carlo sampling. To avoid background pixels in the noisy bounding box representation, they estimate the posteriors in a pixel-wise manner. To boost the pixel-wise posterior estimation, they adopt minibatch Monte Carlo sampling, where only a small portion of pixels are used for inference. Experimental results demonstrate that the proposed visual tracker produces accurate tracking results using a small portion of pixels for the posterior estimation and is comparable to state-of-the-art methods.