Simple paper which revisits the well-known dense connections and feature aggregations typical of architectures like DenseNet and PyramidNet.


They proposed two new architectures : one for classification (fig.3) and one for segmentation (fig.4)

Note that the segmentation architecture is like a U-Net but with conv layers between the encoding and decoding layers. The number of conv layers is inversely proportionnal to the depth of the layers.


Without much surprise, the proposed architectures got quite good results on a variety of classification datasets (like imagenet in Fig.5) and segmentation datasets (like Cityscapes in table 4).