Algorithm 1 Mini-batch gradient ascent.

Input:  The number of employed features F, the training feature vector $f_i\in {\mathbb{R}}^{N\times 1}$, the training ground-truth depth vector $dr\in {\mathbb{R}}^{N\times 1}$, the number of epochs $EP$, the batch size $BS$, and the learning rate $\rho$ Output:  The estimates of ${\theta }_0$, ${\theta }_i$, and ${\sigma }^2$ 1: Initialization $ei=0$, $bn=\lceil N/BS\rceil$, ${\theta }_0$ and ${\theta }_i$ are initialized with random values drawn from $N(0,{10}^{-2})$ 2: while $ei<EP$ do 3:  $bi=0$ 4:   while $bi<bn$ do 5:    if $bi<bn-1$ then 6:     ${\sigma }^2=B{S}^{-1}{\sum }_{j=bi\times BS}^{(bi+1)\times BS-1}{\left\{d{r}_j-\left[{\theta }_0+{\sum }_{i=1}^F{\theta }_i{f}_i\left(j\right)\right]\right\}}^2$ 7:     ${\theta }_0={\theta }_0+\rho {\sigma }^{-2}{\sum }_{j=bi\times BS}^{(bi+1)\times BS-1}\left\{d{r}_j-\left[{\theta }_0+{\sum }_{i=1}^F{\theta }_i{f}_i\left(j\right)\right]\right\}$ 8:     ${\theta }_i={\theta }_i+\rho {\sigma }^{-2}{\sum }_{j=bi\times BS}^{(bi+1)\times BS-1}{f}_i\left(j\right)\left\{d{r}_j-\left[{\theta }_0+{\sum }_{i=1}^F{\theta }_i{f}_i\left(j\right)\right]\right\}$ 9:     check for termination 10:    else 11:     ${\sigma }^2={(N-bi\times BS)}^{-1}{\sum }_{j=bi\times BS}^{N-1}{\left\{d{r}_j-\left[{\theta }_0+{\sum }_{i=1}^F{\theta }_i{f}_i\left(j\right)\right]\right\}}^2$ 12:     ${\theta }_0={\theta }_0+\rho {\sigma }^{-2}{\sum }_{j=bi\times BS}^{N-1}\left\{d{r}_j-\left[{\theta }_0+{\sum }_{i=1}^F{\theta }_i{f}_i\left(j\right)\right]\right\}$ 13:     ${\theta }_i={\theta }_i+\rho {\sigma }^{-2}{\sum }_{j=bi\times BS}^{N-1}{f}_i\left(j\right)\left\{d{r}_j-\left[{\theta }_0+{\sum }_{i=1}^F{\theta }_i{f}_i\left(j\right)\right]\right\}$ 14:     check for termination 15:    end if 16:    $bi=bi+1$ 17:   end while 18:   $ei=ei+1$ 19: end while