Sample labeler for crowd counting based on a first neighbor clustering

Abstract

In terms of crowd counting methods, usually a large number of labeled samples were meeded to train a counting model. In practical applications, to solve the problem of high cost and low efficiency of manually labeling samples, a new sample labeling method for crowd counting was proposed to obtain sufficient training samples. Considering the fact that the distribution of video images having the identical or analogous crowd counts is close to each other in the feature space, with the first neighbor clustering, the video images with similar distance in feature space were divided into the same clusters. A small number of key samples with diversity and representativeness from each cluster were selected by using active sampling strategy and then labeled manually. In the end, the number of labeled samples were propagated to the rest unlabeled samples in each cluster to realize the rapid labeling of samples. The results show that the proposed labeling method can effectively reduce the labeling burden in the crowd counting task, which can achieve better prediction accuracy only with 30% of the labeled samples from other crowd counting methods. Besides, the counting models trained by the obtained sample labels have consistent prediction results compared to the counting models trained by truth value.

Keywords

• crowd counting
• sample labeler
• first neighbor clustering
• active learning
• feature space