COMAL: compositional multi-scale feature enhanced learning for crowd counting

Accurately modeling the crowd's head scale variations is an effective way to improve the counting accuracy of the crowd counting methods. Most counting networks apply a multi-branch network structure to obtain different scales of head features. Although they have achieved promising results, they do not perform very well on the extreme scale variation scene due to the limited scale representability. Meanwhile, these methods are prone to recognize background objects as foreground crowds in complex scenes due to the limited context and high-level semantic information. We propose a compositional multi-scale feature enhanced learning approach (COMAL) for crowd counting to handle the above limitations. COMAL enhances the multi-scale feature representations from three aspects: (1) The semantic enhanced module (SEM) is developed for embedding the high-level semantic information to the multi-scale features; (2) The diversity enhanced module (DEM) is proposed to enrich the variety of crowd features' different scales; (3) The context enhanced module (CEM) is designed for strengthening the multi-scale features with more context information. Based on the proposed COMAL, we develop a crowd counting network under the encoder-decoder framework and perform extensive experiments on ShanghaiTech, UCF_CC_50, and UCF-QNRF datasets. Qualitative and quantitive results demonstrate the effectiveness of the proposed COMAL.