e The neck module is a series of co

e The neck module is a series of convolution layers added forbetter adaptation of the specific task. It could be shared between or addedseparately into subsequent parallel branches. Let n be the number of convolutionlayers in the neck module shared by different supervision. Although the design ofcommon components is simple, the backbone and the first n-layer neck modulecan effectively learn the joint discrimination from the full supervision and weaksupervision. The total number of convolution layers in the neck and subsequentbranch is fixed, but the hyper-parameter n ∈ [0, 3] offers greater flexibility tocontrol the information sharing. When n is 0, each downstream branch has itsown neck module. We denote the network and its output up until the neckmodule as