MDCount: A Lightweight Encoder-Decoder Architecture for Resource-Saving Crowd Counting
Xiaolong Meng, Zhengfei Ren · Journal of Physics Conference Series · 2021
Abstract This work put forwards a lightweight encoder-decoder architecture MDCount designed unambiguously for real-time high-precision crowd counting with constricted calculation resources. The lightweight backbone network MobileNetV2 is tailored to decrease parameter numbers in an acceptable accuracy. The encoder-decoder architecture with atrous spatial pyramid pooling modules is proposed to recover the spatial and contextual information at multiple scales. Experiments on realistic and challenging datasets and comparing contemporary approaches demonstrate that our method MDCount accomplishes equivalent performance with smaller computation costs.