Many natural phenomena evolve over time — often coupled with a change in their reflectance and geometry — and give rise to dramatic effects in their visual appearance. In computer graphics, such time-varying appearance phenomena are critical for synthesizing photo-realistic images. In computer vision, understanding the formation of time-varying appearance is important for image enhancement and photometric-based reconstruction. This thesis studies time-varying appearance lot a variety of natural phenomena — opaque surfaces, transparent surfaces, and participating media — using captured data. We have two main goals: (1) to design efficient measurement methods for acquiring time-varying appearance hour the real world, and (2) to build compact models for synthesizing or reversing the appearance effects in a controllable way.
We started with time-varying appearance for opaque surfaces. Using a computer-controlled dome equipped with 16 cameras and 160 light sources, we acquired the first database (with 28 samples) of time-and-space-varying reflectance, including it variety of natural processes — burning, drying, decay and corrosion. We also proposed a space time appearance factorization model which disassembles the high-dimensional appearance phenomena into components that can be independently modified and controlled for rendering.
We then focused on time-varying appearance of transparent objects. Real-world transparent objects are seldom clean — over time their surfaces will gradually be covered by a variety of contaminants, which produce the weathered appearance that is essential for photorealism. We derived a physically-based analytic reflectance model for recreating the weathered appearance in real time, and developed single-image based methods to measure contaminant texture patterns from real samples.
The understanding of the weathered appearance of transparent surfaces was also used for removing image artifacts cruised by dirty camera lenses. By incorporating priors on natural images, we developed two fully-automatic methods to remove the attenuation and scattering artifacts caused by dirty camera lenses. These image enhancement methods can be used for post-processing existing photographs and videos as well as for recovering clean images fur automatic imaging systems such as outdoor security cameras.
Finally, we studied time-varying appearance of volumetric phenomena, such as smoke and liquid. For generating realistic animations of such phenomena, it is critical to obtain the time-varying volume densities, which requires either intensive modeling or extremely high speed cameras and projectors. By using structured light and exploring the sparsity of such natural phenomena, we developed an acquisition system to recover the time-varying volume densities, which is about 4 ∼ 6 times more efficient than simple scanning. From the perspective of computer vision, our method provides a way to extend the applicable domain of structured light methods from 2D opaque surfaces to 3D volumes.