Fourier feature decorrelation based sample attention for dense crowd localization

Abstract

Crowd localization, which prevails to extract the independent individual features, plays an significant role in critical analysis for crowd scene. Dense trivial features of individual targets are frequently susceptible to interference from complex background features, which makes it difficult to obtain satisfactory predictions for individual targets. Aiming at this issue, a Fourier feature decorrelation based sample attention is proposed for dense crowd localization. The correlation between features are decoupled in the Fourier transform domain, which induces the model to focus more on the true correlation between individual target features and labels. From the perspective of Fourier feature correlation between samples, independence test statistic optimization with cross-covariance operator is developed for feature decorrelation within the sample attention framework. The sample attention with global weight learning is iteratively optimized through matching the prediction loss, which can induce model partial out the spurious correlation between target-irrelevant features and labels. Experimental results show that the method proposed in this paper outperforms the current advanced crowd location methods on public dense crowd datasets.