Configurations of the model architectures of VGGNet-16, ResNext-50 (32x4d), and Swin-B.

In the notation [kernel size × kernel size, channels;…] × N, N denotes the number of times a sequence of convolution layers is repeated, with each layer in the sequence defined by its kernel size and output channels. ‘C’ refers to the cardinality (number of groups). ‘concat n × n’ indicates downsampling by concatenating n × n neighboring features. ‘LN’ abbreviates LayerNorm. ‘win. sz. n × n’ refers to a multi-head self-attention module with a window size of n × n. ‘D-d’ signifies a linear layer with an output dimension of D. △ and ⋆ denote Ext-DFs(CP) and Ext-DFs(FC) respectively. The output of the cells where these symbols are located is utilized as the source of deep features.