Scale-Aware Multi-stage Fusion Network for Crowd Counting

Crowd counting has been widely researched and many hopeful results have been obtained recently. Due to the large-scale variation and complex background noise, accurate crowd counting is still very difficult. In this paper, we raise a simple but efficient network named SMFNet, which focuses on dealing with the above two problems of highly congested noisy scenes. SMFNet consists of two main components: multi-scale dilated convolution block (MDCB) for multi-scale features extraction and U-shape fusion structure (UFS) for multi-stage features fusion. MDCB can address the challenge of scale variation via capturing multi-scale features. UFS provides an effective structure that continuously combines outputs of different stages to achieve the capability of optimizing multi-scale features and increasing resistance to background noise. Compared with the existing methods, SMFNet achieves better performance in capturing effective and richer multi-scale features through progressively multi-stage fusion. To evaluate our method, we have demonstrated it on three popular crowd counting datasets (ShanghaiTech, UCF_CC_50, UCF-QNRF). Experimental results indicate that SMFNet can achieve state-of-the-art results on highly congested scenes datasets.