Count-guided Weakly Supervised Localization Based on Density Map
Weakly supervised localization (WSL) aims at training a model to find the positions of objects by providing it with only abstract labels. For most of the existing WSL methods, the labels are the class of the main object in an image. In this paper, we generalize WSL to counting machines that apply convolutional neural networks (CNN) and density maps for counting. We show that given only ground-truth count numbers, the density map as a hidden layer can be trained for localizing objects and detecting features. Convolution and pooling are the two major building blocks of CNNs. This paper discusses their impacts on an end-to-end WSL network. The learned features in a density map present in the form of dots. In order to make these features interpretable for human beings, this paper proposes a Gini impurity penalty to regularize the density map. Furthermore, it will be shown that this regularization is similar to the variational term of the $\beta$-variational autoencoder. The details of this algorithm are demonstrated through a simple bubble counting task. Finally, the proposed methods are applied to the widely used crowd counting dataset the Mall to learn discriminative features of human figures.
PDF Abstract
Mall
