Towards using count-level weak supervision for crowd counting

作者：

Highlights：

• A count-level weak supervision framework is proposed in reducing the annotation cost for crowd counting.

• The Multiple Auxiliary Task Training (MATT) scheme is introduced to train a better weakly-supervised crowd counter.

• A newly introduced dataset, namely MSCC is designed to evaluate the weakly-supervised crowed counters.

• A superior performance than the straightforward weakly-supervised crowd counting method is achieved.

摘要

•A count-level weak supervision framework is proposed in reducing the annotation cost for crowd counting.•The Multiple Auxiliary Task Training (MATT) scheme is introduced to train a better weakly-supervised crowd counter.•A newly introduced dataset, namely MSCC is designed to evaluate the weakly-supervised crowed counters.•A superior performance than the straightforward weakly-supervised crowd counting method is achieved.

论文关键词：Crowd counting,Count-level annotation,Weak supervision,Auxiliary tasks learning,Asymmetry training

论文评审过程：Received 3 January 2020, Revised 26 July 2020, Accepted 24 August 2020, Available online 25 August 2020, Version of Record 28 August 2020.

论文官网地址：https://doi.org/10.1016/j.patcog.2020.107616