Algorithm 3: The Difference-of-Gaussian filter to invalidate all image pixels belonging to weakly textured areas.
	Data: unfiltered depth, normal and confidence maps ($\mathcal{D}$, $\mathcal{N}$ and $\mathcal{C}$) as well as corresponding reference image ${\mathcal{I}}_{\mathrm{ref}}$.
	Result:  filtered $\mathcal{D}$, $\mathcal{N}$ and $\mathcal{C}$, in which all estimates corresponding to weakly-textured areas in ${\mathcal{I}}_{\mathrm{ref}}$ are removed.
1	Use a Gaussian filter with a kernel of $7\times 7$ pixels to smooth the reference frame ${\mathcal{I}}_{\mathrm{ref}}$, yielding ${\mathcal{I}}_{\mathrm{ref}}^{\mathrm{smooth}}$.
2	Compute the DoG image depicting local image gradients, according to: ${\mathcal{I}}_{\mathrm{ref}}^{\mathrm{DoG}}={\mathcal{I}}_{\mathrm{ref}}-{\mathcal{I}}_{\mathrm{ref}}^{\mathrm{smooth}}$.
3	Apply a binary threshold to compute the DoG mask ${\mathcal{M}}^{\mathrm{DoG}}$, marking all image areas in which the intensity change is greater than $0.5$.
4	Remove activation areas smaller than 7 pixels in ${\mathcal{M}}^{\mathrm{DoG}}$ by applying a speckle filter.
5	Dilate ${\mathcal{M}}^{\mathrm{DoG}}$ with a kernel size of $3\times 3$ pixels to fill small holes in activation areas.
6	Remove deactivation areas smaller than 21 pixels by applying a speckle filter to the inverted DoG mask ${\mathcal{M}}^{\mathrm{D}\mathrm{o}\mathrm{G}\text{-}\mathrm{i}\mathrm{n}\mathrm{v}}=1-{\mathcal{M}}^{\mathrm{DoG}}$.
7	Invalidate pixels in $\mathcal{D}$, $\mathcal{N}$ and $\mathcal{C}$ for which ${\mathcal{M}}^{\mathrm{DoG}}=1$.