Are attention, larger receptive filed and multi task loss good for crowd counting?

In this paper, we transfer some techniques used in computer vision to crowd counting and verify whether they are valid. These techniques are attention mechanism, using larger receptive filed and combination of two different ways ——MSE loss and Bayesian loss. Experiments show that channel attention and spatial attention is not helpful for crowd counting; using larger receptive filed can improve model performance and context-aware method is better than SPP block; combination of two different ways is also not helpful for crowd counting.