Abstract
In this paper, we propose a conceptually simple, advanced and effective head detection framework based on convolutional network. To robustly detect the smaller size of the head in crowded scenes, we propose a new feature extraction strategy which uses a top-down structure and uses lateral connection to combine hierarchical features. Moreover, multi-scale RPN and weight sensitive layer are also explored without increase in the computation costs, as that can reinforce feature representation which is important for identifying small objects. Furthermore, in order to adapt to the needs of the actual application scenarios, we design a model whose size is reduced from 520 M to only 12 M and modify the classification network, which perfect realization of the low calculation and light-weight. We validated our approach on the Brainwash dataset where we show an admirable result compare to the state-of-the-art head detection.