Bias in rendering
What does "unbiased" mean?
Unbiased and consistent are two terms people use to describe error in a rendering algorithm. Although the exact behavior of error may not appear interesting, it has a large impact on how easy the method is to use and how robust it is (in other words, how well it works with a large variety of scenes).
Some people prefer unbiased methods because they make it easy to put a bound on the error. Having a good estimate of the error makes it easier to know when more computation is needed. It can also mean that fewer parameters need to be adjusted in order to produce a good-looking image.
On the other hand, biased methods tend to be much more efficient for common types of scenes than are unbiased methods. These methods make simplifying assumptions which improve efficiency, e.g., that indirect illumination has a low frequency. However, since there are generally no good error bounds for unbiased methods it is difficult to pick the appropriate parameters for a particular rendering.
In general there is a tradeoff between fast, biased methods and robust, unbiased methods. However, there is no hard evidence that unbiased algorithms must be slow or that biased algorithms cannot be robust. The vast majority of commercial applications use fast approximations (rasterization and REYES) which are neither unbiased nor consistent, nor robust!
Unbiased vs. Consistent
People often confuse the term unbiased with the term consistent; for example, you may hear someone say (incorrectly), "photon mapping is unbiased since it converges to the correct solution." These two terms have two precise and different meanings, and it is important to understand the difference.
Consistency is easy to understand: if an image produced by a rendering method approaches the correct solution as some parameter is increased (e.g., number of photons in a photon map), then it is consistent. However, merely knowing that a method is consistent tells you very little. For instance, it does not tell you how quickly the method converges to the correct solution, nor does it give any bound on the error. In general, an estimator FN for a quantity I is consistent for ε if
In other words, as we take more samples the probability of the error being greater than some fixed value ε approaches zero. Most often, "consistent" just means ε = 0, i.e., the estimator approaches the exact answer.
Bias is slightly subtler: a rendering algorithm is unbiased if it produces the correct image on average. If, for example, we averaged many images of the same scene generated by an unbiased rendering algorithm (using different random numbers each time), the result would approach the correct answer. Formally, an estimator is unbiased if
- E[FN − I] = 0,
that is, the expected error or bias is equal to zero — regardless of the value of N. An image path traced with only one sample per pixel is still an unbiased estimate of the correct image. For an unbiased algorithm, the error is proportional to the variance of the estimator, which is equal to , where f is the integrand.
What are common sources of bias in rendering algorithms?
There are many potential sources for bias. In general, a rendering algorithm is biased if it ignores some type of lighting effect, misrepresents the contribution of various lighting effects to the image, or inaccurately computes some quantity of light. Many methods are biased because they ignore certain classes of light paths (e.g., light which bounces off mirrors or is focused through glass). In this case, the result is darker than the correct image, and may lack important visual cues. Other methods introduce bias by interpolating among a sparse set of sample values, ignoring high-frequency features in the integrand and giving the image a blurry appearance. Artifacts such as geometric aliasing (e.g., approximating a smooth surface with triangles) might be considered bias, but we generally assume that the scene described is the one we want to render, and focus on the bias due to the way light transport is handled, i.e., how light is scattered and propagated through the scene.
Which rendering algorithms are consistent? Which are unbiased?
There are very few rendering algorithms which are completely unbiased: even path tracing, which is generally a robust algorithm, excludes several types of light paths. Below are some of the most popular rendering algorithms and their respective sources of error. Descriptions of various types of light paths are given using Heckbert's regular expression notation.
- radiosity - biased and not consistent. Standard radiosity takes into account only paths of the form LD * E -- light bounced diffusely from one surface to another until it hits the eye. Additionally, irradiance is computed over a coarse grid which does not properly capture occlusion (shadows). Although the latter source of error disappears as grid resolution increases, radiosity will always neglect certain types of light paths. Therefore, it is not consistent.
- path tracing essentially unbiased and consistent. Path tracing is usually an accurate way to compute a reference image, but there are a few pathological cases where it will fail. (See , pp 237-240 for further discussion.)
- bidirectional path tracing - unbiased and consistent. By connecting subpaths starting from both the eye and the light, bidirectional path tracing avoids the cases which prevent standard path tracing from being unbiased.
- Metropolis light transport - unbiased and consistent. Since MLT uses an ergodic set of mutation rules (and because care is taken to avoid startup bias), it is unbiased. In other words, as long as we're able to explore all paths which could potentially carry light, detailed balance guarantees that we will converge to the correct answer.
- photon mapping - biased and consistent. There are several sources of bias in photon mapping, but to see that it is biased simply consider what happens when a large number of images generated by a photon mapper are averaged. For example, if we have too few photons in the caustic map, caustics appear blurry due to interpolation. Averaging a large number of blurry caustics will not result in a sharp caustic -- in other words, we don't expect to get the correct answer on average. Some sources of bias are more subtle. For instance, if we have photons deposited in a piece of cloth with many narrow folds, these folds will appear brighter than they should since the density estimate does not take occlusion into account. On the other hand, as we increase the number of photons in the photon map, the region used for each density estimate shrinks to a point. In the limit, a photon used to estimate illumination at a point will correspond to the end of a light subpath at that point. Therefore, as long as the photon map contains a proper distribution of paths, photon mapping is consistent.
- rasterization - biased and not consistent. Most real-time rendering is done using an API such as OpenGL or DirectX which performs all shading based on the information passed to each triangle by a rasterizer. In the past, this framework was limited to integration of paths of the form LDE -- paths which bounce off a single diffuse surface before hitting the eye. It also assumes that any such path is unoccluded, giving an artificially bright result with no shadows. Recently, programmable graphics cards have enabled a larger number of effects, but these tend to be gross approximations to true light transport. Therefore, most images produced via rasterization will not produce the correct solution, regardless of the level of quality. Modern real-time APIs are a mutant example of the tradeoff between speed and robustness: nearly any effect can be achieved in real-time, at the cost of highly special-purpose algorithms.
Will an unbiased algorithm produce a correct image?
Consider a scene where we have an immensely powerful light source in one room and a camera in an adjacent room, where the two rooms are connected only by a microscopic hole (as pictured below). Even with the most efficient rendering algorithm, it may take hundreds of thousands of samples to find a path which connects the light to the camera. If our stopping criterion is a function of our sample variance, we will stop long before we find such a path, since we cannot accurately estimate the constant in our error term. However, we can always make the light bright enough that it will have a substantial effect on the visible illumination. In general, guaranteeing that an image is correct is more difficult than simply finding an unbiased algorithm.