Basis image decomposition of outdoor time-lapse videos
The dominate light in outdoor scenes is the sunlight and
the skylight, and the skylight is distributed on the sky dome.
Since both the sunlight and skylight are far away from the
scenes, it is reasonable to assume that the lighting condition
are the same for each object in the view. We take no account
on any other artificial light sources.
3 Previous model
The global illumination of any scenes can be modeled by
Kajiya’s rendering equation [5]. Radiance of an object is
recorded as RGB value of pixels by CCD of a camera [4],
which are modulated by scale. Therefore, the following
analysis are performed in radiance space.
We adopt the following rendering equations:
L(x, ω, λ, t) = E(x,ω,λ,t)
+
Ω
ρ(x,ω
,ω,λ)S(x,ω
)L(x, ω
,λ,t)
×cos θ(x,ω
)dω
, (1)
where x is an observed point, λ denotes the wavelength of
the light, t is time, ω and ω
are reflection and incident ray
direction, respectively, E(x,ω,λ,t) is the emission light in-
tensity, Ω is the sphere or hemisphere of the incident light,
and ρ(x,ω
,ω,λ) is the BRDF at point x from direction
ω
to ω, and S(x,ω
) is occlusion function, L(x, ω
,λ,t) is
the incident light at x from direction ω
, θ(x,ω
) is angle
between the incident light and the normal at point x.This
equation expresses multiple interreflection among surfaces
and the light arrived at the camera may be reflected arbitrary
times in the scenes.
Liu et al. [11] once proposed a decomposition model of
basis images, by ignoring the interreflection among objects,
and assuming that the sky light is uniformly distributed with
linear HDR function; it arrives at a linear decomposition
equation as follows:
I =E ·C (2)
where I stands for the input images captured with a fixed
camera, E stands for the sunlight and the skylight radiance,
and C is the images of the scene accounting for only the
direct illumination of sunlight and skylight and they are de-
fined as following:
C
sun
(x,λ,t)=
Ω
sun
ρ(x,ω
,ω,λ)S(x,ω
)
×cos θ(x,ω
)dω
,
C
sky
(x,λ,t)=
Ω
sky
ρ(x,ω
,ω,λ)S(x,ω
)
×cos θ(x,ω
)dω
,
(3)
where ω is removed from C
sun
and C
sky
due to fixed cam-
era and geometry in the scene. This model is unsuitable for
the cases of interreflection and non-diffuse surface. In this
paper, we derive the decomposition model in another way.
4 Linear decomposition model
4.1 Global illumination model with interreflection
Based on our assumption, we have E(x) =0 for all pixels of
object surface. A global illumination model should account
for all levels of interreflectance in the scene. We will show
that the linearity is kept for the global illumination model
if the radiance distribution of an area light source is up to
scale.
The rendering equation evaluates the interreflection
through recursive light transfer. An equivalent simulation
is to classify the rays according to the levels of inter-
reflectance. The first level of illumination is by the direct
light source, and then those illuminated surfaces become
light sources of the second level accounting for interreflec-
tion, and so on. Different levels of interreflections proceed
infinitely and they are accumulated as follows:
L(x, ω, λ, t) =
∞
k=0
L
(k)
(x,ω,λ,t) (4)
where k denotes the level of the inter-reflection of light,
L
sun
(x,ω,λ,t) and L
sky
(x,ω,λ,t) integrate all levels of
the inter-reflection of light source. Specifically, according to
rendering equation for each layer of interreflection, we have:
L
(0)
(x,ω,λ,t)=E(x,ω,λ,t),
...
L
(k+1)
(x,ω,λ,t)=
Ω
ρ(x,ω
,ω,λ)S(x,ω
)
×L
(k)
(x, ω
,λ,t)cos θ(x,ω
)dω
.
(5)
From these equations, we can derive the linear property of
decomposition equation.
4.2 Linearity of global illumination
The input images of an outdoor scene are captured at a fixed
viewpoint. For the moment, we assume that the scene is
static, the reflectance of scene points is not limited to Lam-
bertian, but the irradiance of any scene point is entirely due
to natural light.
We assume that the geometry of the scene, the camera po-
sition and orientation, the distribution of the incident light,