Foreground Mask Guided Network for Crowd Counting

Crowd counting in unconstrained scenes is a challenging task due to large scale variations, complex background clutters and severe occlusions, etc. The performance of current networks utilizing multi-path based architectures for better multi-scale representation is constrained by the number of paths. In many cases, existing methods suffer from false responses background such as buildings and trees in complex scenes. To address these issues, we propose an end-to-end network, called Foreground Mask Guided Network (FMGNet), for high-quality density map generation and accurate crowd counting. By employing deep fusion in intermediate layers, the proposed network aggregates multi-scale features in a more efficient way. Moreover, foreground mask features representing the differences between crowd foreground and background are used as guidance information to suppress false responses in background. Extensive experiments on three challenging benchmarks have well demonstrated the effectiveness of the proposed method as well as the superior performance over prior state-of-the-arts.