GTL-ASENet: global to local adaptive spatial encoder network for crowd counting

Crowd counting from a single image is a challenging task due to perspective distortion and large-scale variation in crowd scenes. Many Researches only focus on local features to create density maps which is not effective in handing the challenges. This paper proposes a novel network named global-to-local adaptive spatial encoder network, which focuses on global features to generate a total structure density map of the population distribution, and then utilizes local features to reconstruct the total structure density map in detail to generate high-quality density map. To capture global features, local information and correlate them, we design a contextual module using different kernels with convolution and transposed convolution. To create a density map from global structure to local detail, two branches are designed, the global distribution branch and the local detail branch. The former aims to capture the population distribution region of interest in terms of global structure, and the latter aims to focus on the local details of each unit. Furthermore, to overcome the problem of pixel-wise loss of MSE, this paper proposes an efficient loss function that focuses on perceiving the possible crowd distribution over the whole image. We also apply a new upsampling mechanism that learns to create high-quality density maps on its own is advisable. The proposed network can capture the characteristics of pedestrian distribution and predict accurate results. It is evaluated on four crowd counting datasets (ShanghaiTech, NWPU, UCF_QNRF, UCF_CC_50), it obtains MAE of 67.1 and MSE, and achieves 108.8 in ShanghaiTech and gets MAE of 139.2 and the best MSE of 217.7 in UCF_CC_50 dataset and so on, and our method shows state-of-the-art on all the datasets.