基于自适应策略的人群密度图纠正算法

Density map refinement algorithm based on adaptive strategy

在人群计数任务中,通常使用高斯核平滑标记点图生成的密度图作为中间训练产物。然而,通过对不同密度标签生成方式对训练结果的对比实验,发现并验证生成标签与实际情况存在严重偏差。同时,尽管密度图的生成方式不断被改进,但带来的训练收益十分有限。现有研究主要集中在优化网络结构和损失函数上,却忽视了密度图偏差的校正。受自适应训练算法的启发,设计了一种自适应密度图纠正算法,该算法动态校准生成的密度图分布,并通过模型预测来减少真实标记的偏差。这种方法可以融合到几乎所有的基于卷积神经网络(CNN)或自注意力模型(Transformer)的人群计数模型中,显著提高了模型的准确性和稳健性。在多个人群数据集上的实验结果证明,融合该方法的计数模型实现了更高的准确度和更强的稳健性。

In crowd counting tasks,density maps generated from Gaussian kernel-smoothed point graphs are typically used as intermediate training products.However,comparative experiments on different density label generation methods have revealed a significant discrepancy between generated labels and actual scenarios.Despite ongoing improvements in density map generation,the resulting training gains are limited.Current research focuses primarily on optimizing network structures and loss functions,yet overlooks the correction of density map biases.Inspired by self-adaptive training algorithms,an adaptive density map correction al-gorithm is designed,which dynamically calibrates the distribution of generated density maps,reducing bias in ground-truth through model predictions.This method can be integrated into nearly all crowd counting models based on Convolutional Neural Networks(CNNs)or self-attention models(Transformers),substantially enhancing model accuracy and robustness.Experimental results on various crowd datasets demonstrate that models incorporating this approach achieve higher accuracy and enhanced robustness.