Counting Varying Density Crowds Through Density Guided Adaptive Selection CNN and Transformer Estimation

In real-world crowd counting applications, the densities in an image vary greatly. When facing density variation, humans tend to locate and count targets low-density regions, reason number high-density regions. We observe that CNN focus on local information correlation using a fixed-size convolution kernel Transformer could effectively extract semantic by global self-attention mechanism. Thus, estimate crowds accurately while it is hard properly perceive On contrary, has high reliability but fails sparse Neither nor can well deal with this kind of variation. To address problem, we propose Adaptive Selection Network (CTASNet) which adaptively select appropriate branch for different Firstly, CTASNet generates prediction results Transformer. Then, considering CNN/Transformer low/high-density guided adaptive selection module designed automatically combine predictions Moreover, reduce influences annotation noise, introduce Correntropy based optimal transport loss. Extensive experiments four challenging datasets have validated proposed method.