.png){kind=icon}
.png){kind=icon}
Return
Neural Differential Appearance Equations
.png){kind=icon}
OA
AI
DOI：10.1145/3687900.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
We propose a method to reproduce dynamic appearance textures with space-stationary but time-varying visual statistics. While most previous workdecomposes dynamic textures into static appearance and motion, we focuson dynamic appearance that results not from motion but variations of fun-damental properties, such as rusting, decaying, melting, and weathering. Tothis end, we adopt the neural ordinary differential equation (ODE) to learnthe underlying dynamics of appearance from a target exemplar. We simulatetheODEin two phases. At the warm-up phase, theODEdiffuses a randomnoise to an initial state. We then constrain the further evolution of thisODEto replicate the evolution of visual feature statistics in the exemplar duringthe generation phase. The particular innovation of this work is the neuralODEachieving both denoising and evolution for dynamics synthesis, witha proposed temporal training scheme. We study both relightable (BRDF) nd non-relightable (RGB) appearance models. For both we introduce newpilot datasets, allowing, for the first time, to study such phenomena: ForRGB we provide 22 dynamic textures acquired from free online sources; ForBRDFs, we further acquire a dataset of 21 flash-lit videos of time-varyingmaterials, enabled by a simple-to-construct setup. Our experiments showthat our method consistently yields realistic and coherent results, whereasprior works falter under pronounced temporal appearance variations. Auser study confirms our approach is preferred to previous work for suchexemplars.
Keywords:
Material Appearance
Dynamic Texture
Dynamic (sv)BRDF
Neural Differential Equation
Video
Journal
IF:
9.5
Papers:
4.7K
Citations:
3.6W