Rethinking Spatial Invariance of Convolutional Networks for Object Counting