Pyramid pooling module (PPM)

The key role of Pyramid Pooling Module (PPM) is to generate ensemble high-level feature maps, which represent global context information of multi-scales. In contrast to the SPP that passes the flattened and concatenated multi-level features into FC-NN in the classification tasks, PPM could reduce the loss of information between various sub-levels and extract effective hierarchical global representations. Figure 2b illustrates the configuration of the PPM. The PPM starts with subsampling the convolved features into four parallel pyramid levels with different scales. Larger pooling factor produces coarser features (i.e., similar to global average pooling), while the finer representations are extracted with smaller pooling factors. Then, the bottleneck layer that uses 1 × 1 convolution is applied directly after each pooled features to improve the computation capability by reducing the context dimension to 1/N, where N indicates the pyramid’s level size. For instance, if the level size of the pooling pyramid N=4 (as the case of this work), then the feature maps of each level will be reduced by the factor of 1/4. To get back to the original feature maps right before the pyramid pooling, up-sampling via bilinear interpolation is applied to each pyramid level. Eventually, concatenation of all the up-sampled feature maps with the original feature map is conducted to fuse global context features.

The above two key modules have been exploited to design our proposed network as explained in the following section.