Algorithm 1. Training algorithm.

Require: $x_i$ = input images

Require: L = set of training input indices with known labels

Require: $y_i$ = labels for labeled inputs, i ∈ L

Require:$b{b}_i$ = bounding box of sample i

Require:W_u = unsupervised loss weight

Require:$f_{\theta }\left(x\right)$ = neural network with trainable parameters $\theta$ as student model

Require:$f_{{\theta }^{\prime }}\left(x\right)$ = neural network as teacher model whose parameter ${\theta }^{\prime }$ with initial value $\theta$

Require:$\alpha$ = moving average momentum for parameters

Require:$\eta$ = random Gaussian noise added to the input

fort in [1, num_iterations] do

draw a mini-batch B from labeled and unlabeled samples randomly

$f_i\leftarrow f_{\theta }\left(x_{i\in B},b{b}_{i\in B},\eta \right)$ evaluate network outputs

$f_i{}^{\prime }\leftarrow f_{{\theta }^{\prime }}\left(x_{i\in B},b{b}_{i\in B},{\eta }^{\prime }\right)$ evaluate network outputs

Find triplets T of components <i, j, k> in B where $y_i=y_j$ and $y_i\ne y_j;y_i$ is the pseudo label given by $f_i^{\prime }$ if $i\notin \left(B\cap L\right)$

$loss\leftarrow -\frac{1}{\left|B\cap L\right|}\sum _{i\in \left(B\cap L\right)}{y}_{i}log{f}_i$

+$W\_u(\frac{1}{\left|\left\{i\right|i\notin \left(B\cap L\right)\}\right|}\sum _{i\notin \left(B\cap L\right)}{\Vert f_i-f_i^{\prime }\Vert }^2)$+$W\_u(\frac{1}{\left|T\right|}\sum _{i,j,k\in B}{l}_G\left(i,j\right)+l_G\left(i,k\right))$

update $\theta$ using ADAM optimizer

update ${\theta }^{\prime }$ by ${\theta }_t^{\prime }=\alpha {\theta }_{t-1}^{\prime }+\left(1-\alpha \right){\theta }_t$

end for

return $\theta$