Algorithm 2: Collaboration of Models.

Input   : Sets of pseudo-labeled images: ${\mathcal{X}}_{N_1}^{\widehat{l}},{\mathcal{X}}_{N_2}^{\widehat{l}}$      Vectors of per-class conf. thr.: ${{\mathcal{V}}_{C_T}}_1,{{\mathcal{V}}_{C_T}}_2$      Amount of images to exchange: n      Image-level confidence threshold control: $\lambda$ Output: New sets of ps.-lab. images: ${\mathcal{X}}_{N_1,new}^{\widehat{l}},{\mathcal{X}}_{N_2,new}^{\widehat{l}}$  // $\mathrm{ClassSort}\left(v\right)$ returns the vector of sorted v indices after  // sorting by v values, so that ${\Delta }_i\left[k\right]$ is a class index.           ${\Delta }_1$ ←$\mathrm{ClassSort}\left({{\mathcal{V}}_{C_T}}_2-{{\mathcal{V}}_{C_T}}_1\right)$           ${\Delta }_2$ ←$\mathrm{ClassSort}\left({{\mathcal{V}}_{C_T}}_1-{{\mathcal{V}}_{C_T}}_2\right)$  // $\mathrm{ClassImageList}\left(\mathcal{X}\right)$ returns a vector so that ${\mathcal{S}}_i\left[k\right]$ is the  // list of images in $\mathcal{X}$ containing pseudo-labels of class k.           ${\mathcal{S}}_1$ ←$\mathrm{ClassImageList}\left({\mathcal{X}}_{N_1}^{\widehat{l}}\right),$           ${\mathcal{S}}_2$ ←$\mathrm{ClassImageList}\left({\mathcal{X}}_{N_2}^{\widehat{l}}\right)$ ${\mathcal{X}}_{N_1,new}^{\widehat{l}},{\mathcal{X}}_{N_2,new}^{\widehat{l}}$ ←$\varnothing ,\varnothing$          $k,N_c$ ←$0,\mathrm{Num.\; Classes}$

return ${\mathcal{X}}_{N_1,new}^{\widehat{l}},{\mathcal{X}}_{N_2,new}^{\widehat{l}}$