This paper extends previous shadow mapping work into dynamic scences under complex lighting such as area light sources or environment lighting.
This extension relies on fast approximate technique for computing soft shadows. This new soft shadow mapping uses two convolution operaton to compute the filtering kernel size and prefiltering shadow map respecitively. The latter convolution is their previous work: convolution shadow mapping(CSM)
In other word, they extend CSM with adaptive filteirng kerenl. Furthermore, the kerenl computation is in constant time with the same mathematical framework.
To render the environement map lighting, the decompose environment map into many area lights. Then they use this high-quality faster soft shadow mapping method to calculatereal-time all-frequency shadows in dynamic scenes.
However there are still exist some limitations:
First they assume that all blocker has he same depth within the convolution kernel, that is:
Sf(x)= Sum(w(q) * f(d(y), z(p-q))
here we think all d(y) equal to d(x). This is a popular problem in most of advanced shadow mapping method
Second, they use Fourier basis function to approximate results. However Fourier basis easily produce some ringing artifacts. In addition
the kernel size computation involves with divison. it is easilly introduce error.
Third. like most of shadow mapping, they don't support complex BRDF effect.
In conclusion, they propose a new shadow mapping method with convolution to make all-frequency shadow in dyanmic scence. With greedy algorithm to generate area light from environment map, they extend previous shadow mapping into complex lighting.