Semi-supervised Crowd Counting via Density Agency

Abstract

In this paper, we propose a new agency-guided semi-supervised counting approach. First, we build a learnable auxiliary structure, namely the density agency to bring the recognized foreground regional features close to corresponding density sub-classes (agents) and push away background ones. Second, we propose a densityguided contrastive learning loss to consolidate the backbone feature extractor. Third, we build a regression head by using a transformer structure to refine the foreground features further. Finally, an efficient noise depression loss is provided to minimize the negative influence of annotation noises. Extensive experiments on four challenging crowd counting datasets demonstrate that our method achieves superior performance to the state-of-the-art semisupervised counting methods by a large margin. Code is available. CCS CONCEPTS• Computing methodologies → Computer vision; Semi-supervised learning settings; Learning from implicit feedback.

Region feature smoothness assumption for weakly semi‐supervised crowd counting

“…The experimental results of UCF-QNRF dataset is shown in Table 2, our method improved MAE by 0.7% and MSE by 23.1% compared with Tran-sCrowd, 14 improved improved MAE by 7.6% and MSE by 3.2% compared with DACount. 22 In summary, our method achieves superior experimental results with a smaller amount of data in all datasets.…”

“…Compared with existing state‐of‐the‐art weakly supervised method, our method achieved competitive results. compared with DACount, 22 with a small amount of labeling data, our method improved MAE by 0.7% and MSE by 23.1%. JHU‐Crowd++ dataset has a very large number of crowd images, since the density of the crowd image is very wide.…”

“…In order to evaluate the effectiveness of our proposed method, we have selected state‐of‐the‐art semi‐supervised and weakly‐supervised crowd counting methods for comparison. Among them, MT, 36 L2R, 11 GP, 12 IRAST, 13 SUA, 37 S$$$ {}^2 $$$ FPR, 10 DACount 22 are semi‐supervised crowd counting methods, TransCrowd, 14 MATT, 17 CCTrans 15 are weakly supervised methods. Comprehensive experiments are performed on five benchmark datasets.…”

“…To be specific, this method calculates the ranking loss within different levels of features and fuses them at the end. Lin et al 22 presented a crowd density feature agency based on different crowd density distribution of regions. This feature agency is constantly updated to distinguish different crowd region densities and environmental backgrounds.…”

Counting animals in aerial images with a crowd counting model

“…IRAST [43] designs a series of interrelated binary classification tasks and exploits the underlying constraints among them. DACount [44] employs multiple density agents to promote similar feature representations for patches with close density values. Distinct to previous methods that fixated on the accuracy of individual patches, our approach is centered on improving the model's holistic understanding of crowd scenes, resulting in outstanding performance compared to existing methods.…”