Algorithm 1 The algorithm for CS sampling with multiple MRs in the proposed MRCS

Input: A natural image of $W\times H$ with $c(c\ge 1)$ channels, and k measurement matrices such as ${\mathbf{\Phi }}_h$ generated with a higher measurement rate and ${\mathbf{\Phi }}_l$ generated with a lower measurement rate.

Output: The results of half-precision CS measurement $\mathbf{y}$

1:Detect the bounding boxes of n target objects (e.g., persons, bicycles and cars) with the proposed object detector MYOLO3. The bounding box $b_j(j<n)$ is denoted as $[x_j,y_j,w_j,h_j]$, where $x_j$ and $y_j$ declare the coordinates of its top-left corner, $w_j$ and $h_j$ declare its width and height respectively. 2:Divide the input image into patches of size $33\times 33$ without overlap, and get $wp\times hp=ceil\left(\right(W-1)/33+1)\times ceil\left(\right(H-1)/33+1)$ image patches. 3:Declare a $wp\times hp$ matrix $\mathbf{P}=\left\{p_{st}\right\},p_{st}\in \{0,1\}$ as the identifier of the divided image patches. The initial values of the elements of $\mathbf{P}$ are set as 0. 4:while There are detected bounding boxes to be identified in the 1-th channel, i.e., $j<n$ do 5: Compute the starting indexes $s_x$ and $s_y$ for the j-th image patch, i.e., $s_x=ceil(x_j/33)$ and $s_y=ceil(y_j/33)$; 6: Compute the ending indexes $e_x$ and $e_y$ for the j-th image patch, i.e., $e_x=ceil((x_j+w_j)/33)$ and $e_y=ceil((y_j+h_j)/33)$; 7: while $s\ge s_x$ and $s\le e_x$ do 8:  while $t\ge s_y$ and $t\le e_y$ do 9:   $p_{st}=1$; 10:   $t=t+1$; 11:  end while 12:  $s=s+1$; 13: end while 14: j = j+1; 15: end while 16:while There are channels to be sampled, i.e., $i\ge c$ do 17: while $s<wp$ do 18:  while $t<hp$ do 19:   Denote the pixels of the image patch identified with $p_{st}$ in the i-th channel as ${\mathbf{x}}_{ist}$; 20:   if $p_{st}=1$ then 21:    Sample the image patch ${\mathbf{x}}_{ist}$ with ${\mathbf{\Phi }}_h$, i.e., ${\mathbf{y}}_{ihst}={\mathbf{\Phi }}_h{\mathbf{x}}_{ist}$; 22:   else 23:    Sample the image patch ${\mathbf{x}}_{ist}$ with ${\mathbf{\Phi }}_l$, i.e., ${\mathbf{y}}_{ilst}={\mathbf{\Phi }}_l{\mathbf{x}}_{ist}$; 24:   end if 25:   $t=t+1$; 26:  end while 27:  $s=s+1$; 28: end while 29: Concatenate all CS measurements sampled with ${\mathbf{\Phi }}_h$ and ${\mathbf{\Phi }}_l$ respectively, represent the values with 16-bit half-precision floats, and then obtain the half-precision CS measurement ${\mathbf{y}}_i={\left[{\mathbf{y}}_{ih}\right]}_{\mathbf{FP}\mathbf{16}}\cup {\left[{\mathbf{y}}_{il}\right]}_{\mathbf{FP}\mathbf{16}}$ for the i-th channel; 30: $i=i+1$; 31: end while 32:Combine ${\mathbf{y}}_i$ sampled with multiple MRs in each channel and then obtain the final half-precision CS measurement $\mathbf{y}$.