Uncertainty Estimation and Sample Selection for Crowd Counting

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