Cross-head Supervision for Crowd Counting with Noisy Annotations

Mingliang DaiZhizhong HuangJiaqi GaoHongming ShanJunping Zhang

Mar 2023

摘要

Noisy annotations such as missing annotations and location shifts often existin crowd counting datasets due to multi-scale head sizes, high occlusion, etc.These noisy annotations severely affect the model training, especially fordensity map-based methods. To alleviate the negative impact of noisyannotations, we propose a novel crowd counting model with one convolution headand one transformer head, in which these two heads can supervise each other innoisy areas, called Cross-Head Supervision. The resultant model, CHS-Net, cansynergize different types of inductive biases for better counting. In addition,we develop a progressive cross-head supervision learning strategy to stabilizethe training process and provide more reliable supervision. Extensiveexperimental results on ShanghaiTech and QNRF datasets demonstrate superiorperformance over state-of-the-art methods. Code is available athttps://github.com/RaccoonDML/CHSNet.

展开全部

图表提取