BlurPool inserts a fixed low-pass filter before each downsampling operation to first smooth the feature map and then downsample, decomposing traditional pooling into first performing max value selection, then smoothing through a fixed blur kernel, and finally stride sampling.