I = argmin
¯
I
(||α(
¯
I − I
g
)||
p
+ ||
H
dx
¯
I
H
dy
¯
I
H
dt
¯
I
H
dxdt
¯
I
H
dytdt
¯
I
−
I
dx
I
dy
I
dt
I
dxdt
I
dydt
||
p
) (25)
Where H
∗
denotes the finite difference operator and I
g
is the sequence of primal images. This
is basically identical to the reconstruction in Equation (11) except for the fact that we reconstruct a
sequence of images. To achieve arbitrary length reconstruction [Man+16] use a sliding window of 20
frames.
8 Discussion
In this report we presented the state of the art in gradient domain rendering. We have provided an
analysis that shows, under certain simplifications, how the process of sampling gradients followed by
Poisson reconstruction suppresses high frequency noise and reduces the overall error. The papers which
describe the individual techniques already have detailed analysis comparing their work to the non-gradient
domain counterparts. Also, a direct comparison between gradient domain methods is not very insightful.
For example, it is obvious that gradient domain path tracing is going to perform worse than gradient
domain bidirectional path tracing in scenes with difficult light transport. Gradient domain rendering does
not change the characteristics of the underlying technique, but provides a constant error reduction factor
(see Equation (20)). Even with a very bad shift mapping, which always produces non-symmetric paths,
we will not do worse than the base algorithm at equal sample counts as long as we handle non-symmetric
shifts correctly. Something that is worth investigating is the run time overhead of gradient sampling,
which is dominated by the time taken to construct offset paths. This is why equal time comparisons tend
to be more meaningful than equal sample count comparisons. Additionally, it would be interesting to get
statistics about how scene geometry influences the effectiveness of shift mappings. Getting insights into
where the problem areas lie is helpful in the construction of new shift mappings.
8.1 Equal Time Comparisons
We want to compare the effectiveness of gradient sampling in various scenarios. For this we measure
the relative mean squared error (relMSE) of the generated images after letting them render for fixed
amount of time. The relMSE of a sampled image I is defined as
rel
= ||I
ref
− I||
2
/||I
ref
||
2
where
I
ref
is a converged reference image, usually obtained via path tracing a large number of samples. For
each technique we render two images, one using the classical version of the techniques and one using
gradient sampling. We measure the relative improvement by dividing their relMSEs. We also compute
the predicted improvement by computing the factor given by Equation (20) for the reference image.
Figure 18 shows the result of this investigation. We use three scenes with different light transport
characteristics. The scene a) has little power in the gradients but highly specular materials. We predict
an improvement do to gradient sampling by 28x over classical sampling, but neither G-PT nor G-BDPT
achieve factors close to this. Scene b) is the classic ”Sponza” scene. It contains only diffuse materials
and is illuminated by a large light, ideal conditions for both path tracing as well as the shift mapping
employed for G-PT. In fact, we see that G-PT almost achieves the ideal improvement over PT. On the
other hand, G-BDPT only reduces relMSE by 8 times, half of the theoretically predicted value. At last,
the scene c) presents a difficult light transport scenario for PT. This is the only scene were G-BDPT
outperforms PT in terms of relative error reduction.
8.2 Analysing Shift Mappings
To evaluate the performance of shift mappings we compute the percentage of failed paths per pixel. For
this we modify the implementation by [Ket+15] and add an additional output. In Figure 19 we display
the final image on the left and a heat map of failed samples on the right. Generally, we can see that shift
failures concentrate on edges, though the specularity of the interacting materials also seems to have an
influence.
To further investigate this we study a scene with materials of varying reflectivity and otherwise fairly
simple light transport in Figure 20. Of interest here is that the shift does not fail frequently when hitting