Scale-Recursive Network with point supervision for crowd scene analysis

Crowd scene analysis, and in particular its density estimation, is a challenging task due to the lack of spatial information, scale variation, and the large amount of supervised-learning parameters. In order to address these challenges, we propose a Scale-Recursive encoder-decoder Network with Point Supervision (SRN+PS). On the one hand, an encoder-decoder recurrent structure uses features between adjacent scales to tackle scale variation, and a novel loss function, called the row vector-based counting loss, is proposed to focus on the crowd counting accuracy. On the other hand, we employ an additional point segmentation task in training and combine features learned from the two tasks above. The Euclidean loss, row vector-based counting loss, and two-label focal loss are integrated by a joint training scheme, which improves both the quality of density map estimation and the performance of crowd counting. Finally, we propose a weakly supervised framework based on the SRN structure and the Convolutional Winner-Take-All(CWTA) module. In this framework, most parameters are obtained by unsupervised learning with the exception of a few which are tuned by supervised learning in model training. As a result, our multi-scale structure can obtain salient object sparse spatial features from unsupervised learning. Experiments on the ShanghaiTech, UCF_CC_50 and UCSD datasets demonstrate the effectiveness of our proposed method. (C) 2019 Elsevier B.V. All rights reserved.