Algorithm 1: Superpixel-based projection in 3D
	Input:  Lower-view RGBD image ($I\in {[0,255]}^{H\times W\times 3}$ and $D\in {[0,255]}^{H\times W}$), upper-view 3D semantic reconstruction (including point cloud $P_{upper}\in {\mathbb{R}}^{N\times 3}$ and labels $L_{upper}\in {\{0,1,\cdots ,L\}}^N$, where $L\in \mathbb{N}$ denotes the number of semantic labels). Output:  Lower-view semantic segmentation ($S\in {\{0,1,\cdots ,L\}}^{H\times W}$)
1	Run superpixel segmentation (SLIC) on the RGB image (I) to find coherent image regions ($SP$).
2	Using the depth image (D), project the superpixel segmentation ($SP$) onto 3D to obtain the lower-view point cloud ($P_{lower}$).
3	$\left[\mathrm{Optional}\right]$ Downsample the point cloud ($P_{lower}$).
4	Match the lower-view point cloud ($P_{lower}$) with the upper-view 3D semantic reconstruction ($P_{upper},L_{upper}$) and determine the corresponding semantic label of each superpixel ($L_{lower}$).
5	Project the labels of the lower-view semantic point cloud ($L_{lower}$) onto the image plane to obtain the semantic segmentation (S).