Uncertainty Estimation and Sample Selection for Crowd Counting

TL;DR

This paper introduces a crowd counting method that predicts density maps with uncertainty estimates, enabling efficient domain adaptation with significantly reduced labeled data.

Contribution

It develops a neural network that models crowd density with Gaussian distributions and uses uncertainty-based sample selection for domain adaptation.

Findings

Reduces labeled data needed for domain adaptation to 17% of target domain samples.

Surpasses previous state-of-the-art performance on NWPU and Shanghaitech datasets.

Quantifies prediction uncertainty to improve crowd counting accuracy.

Abstract

We present a method for image-based crowd counting, one that can predict a crowd density map together with the uncertainty values pertaining to the predicted density map. To obtain prediction uncertainty, we model the crowd density values using Gaussian distributions and develop a convolutional neural network architecture to predict these distributions. A key advantage of our method over existing crowd counting methods is its ability to quantify the uncertainty of its predictions. We illustrate the benefits of knowing the prediction uncertainty by developing a method to reduce the human annotation effort needed to adapt counting networks to a new domain. We present sample selection strategies which make use of the density and uncertainty of predictions from the networks trained on one domain to select the informative images from a target domain of interest to acquire human annotation. We show that our sample selection strategy drastically reduces the amount of labeled data from the target domain needed to adapt a counting network trained on a source domain to the target domain. Empirically, the networks trained on UCF-QNRF dataset can be adapted to surpass the performance of the previous state-of-the-art results on NWPU dataset and Shanghaitech dataset using only 17$\%$ of the labeled training samples from the target domain.