Fluid Volume Modeling from Sparse Multi-view Images by Appearance Transfer

摘要

We propose a method of three-dimensional (3D) modeling of volumetricrnfluid phenomena from sparse multi-view images (e.g., onlyrna single-view input or a pair of front- and side-view inputs). Thernvolume determined from such sparse inputs using previous methodsrnappears blurry and unnatural with novel views; however, ourrnmethod preserves the appearance of novel viewing angles by transferringrnthe appearance information from input images to novelrnviewing angles. For appearance information, we use histogramsrnof image intensities and steerable coefficients. We formulate thernvolume modeling as an energy minimization problem with statisticalrnhard constraints, which is solved using an expectation maximizationrn(EM)-like iterative algorithm. Our algorithm begins withrna rough estimate of the initial volume modeled from the input images,rnfollowed by an iterative process whereby we first render thernimages of the current volume with novel viewing angles. Then,rnwe modify the rendered images by transferring the appearance informationrnfrom the input images, and we thereafter model the improvedrnvolume based on the modified images. We iterate these operationsrnuntil the volume converges. We demonstrate our methodrnsuccessfully provides natural-looking volume sequences of fluidsrn(i.e., fire, smoke, explosions, and a water splash) from sparse multiviewrnvideos. To create production-ready fluid animations, we furtherrnpropose a method of rendering and editing fluids using a commerciallyrnavailable fluid simulator.