Abstract—We propose a Convolutional Neural Network for counting objects and persons in images. Utilizing a sequence of images, it is possible to count with respect to a movement direction (e.g. UCSD pedestrian dataset). The proposed Number Convolutional Neural Network (NCNN) directly outputs the desired count and thus does not require additional counting steps or dense maps as intermediate step. It cannot only be superior with respect to the mean absolute error evaluated on already known datasets. It also requires only the count as ground truth data and is thus easily and quickly applied to a variety of new problem statements. Additionally, it is able to count with respect to a movement direction by integrating time-dependent information.

Index Terms—deep learning, CNN, crowd counting

Cite: Oliver Urbann and Jonas Stenzel, "A Convolutional Neural Network that Self-Contained Counts," Journal of Image and Graphics, Vol. 7, No. 4, pp. 112-116, December 2019. doi: 10.18178/joig.7.4.112-116