Consistent-Point: Consistent Pseudo-Points for Semi-Supervised Crowd Counting and Localization

TL;DR

Consistent-Point addresses two forms of pseudo-point inconsistency in semi-supervised, point-localization crowd counting and localization by introducing Position Aggregation (PA) and Instance-wise Uncertainty Calibration (IUC) within a mean-teacher framework built on P2PNet. PA smooths pseudo-point positions by averaging neighboring auxiliary proposal-points, while IUC weights pseudo-points by their classification confidence to stabilize their semantic meaning. Empirically, the approach delivers state-of-the-art localization across five datasets and multiple labeling ratios and also achieves competitive counting results, including surpassing some fully supervised baselines under label scarcity. The work demonstrates that enforcing pseudo-point consistency substantially improves learning signals in semi-supervised crowd counting and localization, with practical benefits for annotation-efficient deployment.

Abstract

Crowd counting and localization are important in applications such as public security and traffic management. Existing methods have achieved impressive results thanks to extensive laborious annotations. This paper propose a novel point-localization-based semi-supervised crowd counting and localization method termed Consistent-Point. We identify and address two inconsistencies of pseudo-points, which have not been adequately explored. To enhance their position consistency, we aggregate the positions of neighboring auxiliary proposal-points. Additionally, an instance-wise uncertainty calibration is proposed to improve the class consistency of pseudo-points. By generating more consistent pseudo-points, Consistent-Point provides more stable supervision to the training process, yielding improved results. Extensive experiments across five widely used datasets and three different labeled ratio settings demonstrate that our method achieves state-of-the-art performance in crowd localization while also attaining impressive crowd counting results. The code will be available.

Figures (4)

• Figure 1: Annotator subjectivity introduces position inconsistency for manual annotation points. The position and classification inconsistency of pseudo-points arise from the regression branch and classification branch of the point-localization teacher model with changing parameters in the training process, respectively.
• Figure 2: The pipeline of Consistent-Point. We design two modules for point-localization-based crowd counting under the mean-teacher paradigm. Position Aggregation (PA) alleviates the position inconsistency from the regression branch; Instance-wise Uncertainty Calibration (IUC) alleviates the class inconsistency arising from the classification branch.
• Figure 3: Some qualitative results of Baseline and our Consistent-Point. Green points represent ground truth points, while red dots indicate the predicted points. The white numbers in the bottom-right corner of the images denote the total count of corresponding points.
• Figure 4: Ablation study on the unlabeled loss weight $\lambda$.