I met this term in the paper "Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition":"3.1.2 Multi-level Pooling Improves Accuracy In Table 2 (b) we show the results using single- size training. The training and testing sizes are both 224×224. In these networks, the convolutional layers have the same structures as the corresponding base- line models, whereas the pooling layer after the final convolutional layer is replaced with the SPP layer. For the results in Table 2, we use a 4-level pyramid. The pyramid is {6×6, 3×3, 2×2, 1×1} (totally 50 bins). For fair comparison, we still use the standard 10- view prediction with each view a 224×224 crop. Our results in Table 2 (b) show considerable improvement over the no-SPP baselines in Table 2 (a). Interestingly, the largest gain of top-1 error (1.65%) is given by the most accurate architecture. Since we are still using the same 10 cropped views as in (a), these gains are solely because of multi-level pooling."
thanks a lot