Fig 4. Pooling module and deconvolution module to effectively increase the number of layer Conv4_3’s feature maps.

Above the dot-dash line was pooling module. We used multiple pooling methods to increase Conv4_3’s feature maps for detection. Below the dot-dash line was deconvolution module. We adopted deconvolution from Conv4_3’s neighbor feature pyramid layer Fc7 to increase Conv4_3’s feature maps for detection.