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

Yuda Zou, Zelong Liu, Yuliang Gu, Bo Du, Yongchao XuarXiv 2025 · DOI · arXiv

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.

Introduction

Crowd counting—identifying the number and location of individuals in a scene—is a critical task for public safety and urban planning. While deep learning has significantly advanced this field, these models typically rely on exhaustive manual annotations, where every head in an image must be clicked. Semi-supervised learning (SSL) offers a promising alternative by leveraging a small set of labeled data alongside a larger pool of unlabeled images. A common approach involves a "Teacher" model generating "pseudo-points" to supervise a "Student" model.

However, a significant challenge in this paradigm is the instability of pseudo-points. As the Teacher model evolves during training, the generated labels often exhibit "jitter" in their spatial coordinates or fluctuate between "person" and "background" classifications. This inconsistency introduces noise into the training signal, limiting the performance of the Student model.

Figure 1: Illustration of position and class inconsistency in pseudo-points compared to manual annotations.

Methodology: The Consistent-Point Framework

To address these challenges, the authors propose Consistent-Point, a framework built upon the Mean-Teacher paradigm using the P2PNet backbone. The core innovation lies in two modules designed to stabilize pseudo-labels:

1. Position Aggregation (PA)

To mitigate spatial "jitter," the PA module moves away from relying on a single predicted point. Instead, it identifies the four nearest "auxiliary" proposal points and computes their average position. This spatial smoothing effectively reduces the variance of the predicted locations. Because localization errors among neighboring points often follow similar distributions, averaging them provides a more stable and reliable target for the Student model.

2. Instance-wise Uncertainty Calibration (IUC)

To handle class inconsistency, the IUC module introduces a weighting mechanism for the loss function. Each pseudo-point is assigned a weight based on its classification confidence score. High-confidence predictions provide strong supervision, whereas ambiguous points—those that frequently flip between classes—are down-weighted. This prevents the model from learning from unreliable or noisy labels.

Figure 2: The Consistent-Point pipeline, showing how PA and IUC modules refine the Teacher's output before supervising the Student.

Experimental Results

The Consistent-Point method was rigorously evaluated on five major benchmarks: ShanghaiTech (Parts A & B), UCF-QNRF, JHU-Crowd++, and NWPU-Crowd.

Performance Highlights

  • State-of-the-Art Accuracy: Consistent-Point achieved SOTA results across all datasets and labeled-data ratios (5%, 10%, and 40%). For instance, on the NWPU-Crowd dataset with only 10% labeled data, it reached an F1-measure of 73.8%.
  • Significant Counting Improvements: On ShanghaiTech Part A (10% labeled), the Mean Absolute Error (MAE) was reduced from a baseline of 81.58 to 63.55.
  • Robustness to Density: In dense regions of the UCF-QNRF dataset, the model using 40% labeled data actually outperformed the fully supervised (100% labeled) baseline. This suggests that the consistent pseudo-labeling process may even correct for human errors or subjectivity inherent in manual annotations of extremely crowded scenes.

Conclusion

Consistent-Point demonstrates that the stability of pseudo-labels is just as important as their accuracy in semi-supervised crowd counting. By introducing Position Aggregation and Instance-wise Uncertainty Calibration, the framework provides a steady and high-quality supervision signal. This allows for high-performance crowd localization and counting even with minimal labeled data, offering a practical solution for real-world applications where manual annotation is prohibitively expensive. Future work will explore dynamic aggregation strategies to better handle extremely high-density scenarios.

Summary updated

BibTeX
@misc{zou2025consistent,
  author = {Yuda Zou and Zelong Liu and Yuliang Gu and Bo Du and Yongchao Xu},
  title = {{Consistent-Point: Consistent Pseudo-Points for Semi-Supervised Crowd Counting and Localization}},
  year = {2025},
  eprint = {2503.12441},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.12441},
}
Figure