Multi-task semi-supervised crowd counting via global to local self-correction

Abstract

In this paper, we propose a novel multi-task semi-supervised method. To sufficiently exploit massive unlabeled data, multi-task pseudo-labels and global to local self-correction strategy are proposed. Specifically, labeled images and massive amounts of unlabeled images with proposed multi-task pseudo-labels are leveraged for model optimization. The density level of the whole image is predicted in classification task. The density is estimated in density regression task. The crowd area is segmented out in segmentation task. To suppress incorrect predictions caused by the inevitable noises from some unlabeled data misleading the model, the counting relationship between classification task and density task is exploited to propose the global self-correction strategy, and the semantic consistency between density task and segmentation task is mined to propose the local self-correction strategy. The classification task and segmentation task contribute in generating the final highly refined density map from the density task. Extensive experiments on six benchmark datasets indicate the superiority of our method over the SOTA methods in semi-supervised paradigm.