MobileCount: An efficient encoder-decoder framework for real-time crowd counting

In this work, we propose a computation-efficient encoder-decoder architecture, named MobileCount, which is specifically designed for high-accuracy real-time crowd counting on mobile or embedded devices with limited computation resources. For the encoder part, MobileNetV2 is tailored in order to sign...

Full description

Saved in:
Bibliographic Details
Published in:Neurocomputing (Amsterdam) Vol. 407; pp. 292 - 299
Main Authors: Wang, Peng, Gao, Chenyu, Wang, Yang, Li, Hui, Gao, Ye
Format: Journal Article
Language:English
Published: Elsevier B.V 24.09.2020
Subjects:
Crowd counting
Fully convolutional networks
Knowledge distillation
Light-weight Neural Networks
Fully convolutional networks
Light-weight Neural Networks
Knowledge distillation
Crowd counting
ISSN:0925-2312, 1872-8286
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this work, we propose a computation-efficient encoder-decoder architecture, named MobileCount, which is specifically designed for high-accuracy real-time crowd counting on mobile or embedded devices with limited computation resources. For the encoder part, MobileNetV2 is tailored in order to significantly reduce FLOPs at a little cost of performance drop, which has 4 bottleneck blocks preceded by a max pooling layer of stride 2. The design of decoder is motivated by Light-weight RefineNet, which further boosts counting performance with only a 10% increase of FLOPs. In comparison with state-of-the-arts, our proposed network is able to achieve comparable counting performance with 1/10 FLOPs on a number of benchmarks. At last, we propose a multi-layer knowledge distillation method to further boost the performance of MobileCount without increasing its FLOPs.
ISSN:0925-2312
1872-8286
DOI:10.1016/j.neucom.2020.05.056