Ground truth intrinsic decompositions

LightField Images

Random filters	Input	Random filter per frame	Our regularization (λ_t=0, λ_s=10)
Intrinsic Reflectance	Input(source)	[Garces et al. 2016]	[Zhao et al.] per frame	Our regularization (λ_t=0, λ_s=1)
Intrinsic Reflectance	Input(source)	[Garces et al. 2016]	[Zhao et al.] per frame	Our regularization (λ_t=0, λ_s=1)

LightField Videos

Intrinsic Reflectance	Input (11 views) (source)	[Bell et al. 2014] per frame	[Lang et al.] regularization per view	[Bonneel et al. 2015] regularization per view	Our regularization (λ_t=0.1, λ_s=0.1)	Our regularization, filter transfer, subsample factor 10x
Intrinsic Shading	Same input	[Bell et al. 2014] per frame	[Bonneel et al. 2015] regularization per view	[Bonneel et al. 2015] regularization scanning all views (more details) -- includes filter transfer with 10x subsampling	Our regularization (λ_t=0.1, λ_s=0.1)	Our regularization, filter transfer, subsample factor 10x
Tone Mapping	Same input	[Durand and Dorsey, 2012] per frame	Our regularization	Our regularization (λ_t=0.1, λ_s=0.1), filter transfer, subsample factor 10x
Neural style transfer	Same input (2 views only)	[Gatys et al.] per frame	Our regularization
Deep Dream	Same input (2 views only)	Deep Dream per frame	Our regularization (4 iterations)
Intrinsic Reflectance	Input (10 views)	[Bell et al. 2014] per frame	Filter transfer experiment: Anchor frames sub-sampled instead of fully processed (3 views) (λ_t=0.2, λ_s=0.02)	Filter transfer with anchor frames fully processed (λ_t=0.2, λ_s=0.05)
Neural Style Transfer	Same input (8 views only, downsampled)	[Gatys et al.] per frame	Our regularization after 2 iterations	Our regularization after 3 iterations	Our regularization after 4 iterations (final)
Deep Dream	Same input (2 views only, downsampled)	Deep Dream per frame	Our regularization (4 iterations)
Intrinsic Reflectance	Input (7 views)	[Bell et al. 2014] per frame	Our regularization (λ_t=0.2, λ_s=0.02)
Intrinsic Shading	Same input	[Bell et al. 2014] per frame	Our regularization
Photoshop Oil Paint	Input (7 views)	Per frame Oil Paint	Our regularization -- failure case (λ_t=0.1, λ_s=0.1)
Photoshop Cutout	Same input	Per frame cut out	Our regularization

Multi-view videos

Color2gray	Input (6 views)	[Lu et al. 2012] per frame	Our regularization (λ_t=0.2, λ_s=0.2)	Experiment: using PCA Flow instead of modified PatchMatch (λ_t=0.2, λ_s=0.2)
Adobe Premiere AutoColor	Input (5 views)	Premiere Auto Color (no smoothing)	Our regularization (λ_t=0.02, λ_s=0.02), filter transfer, subsample factor 2x	Premiere Auto Color (with smoothing)

Stereo videos

Adobe Photoshop AutoColor	Input (2 views)	Photoshop Autocolor per frame	Our regularization (λ_t=0.01, λ_s=0.01)
Adobe Photoshop Dehazing	Input (2 views)	Photoshop Dehazing per frame	Our regularization (λ_t=0.05, λ_s=0.05)