Research On BRDF Representation Based On Neural Processes

Measured BRDFs faithfully record materials' reflectance profiles and can produce renderings with a remarkable level of fidelity.As a result,they are used in many realworld applications,where visual realism is crucial.In this thesis,a compact low-dimensional representation is obtained for measured BRDFs by leveraging network-based method,Neural Processes(NPs).Unlike prior methods that treat those BRDFs as discrete high-dimensional matrices or tensors,our technique considers measured BRDFs as continuous functions and projects it to the function latent spaces.Specifically,provided a collection of BRDFs,such as materials in MERL and EPFL datasets,our method learns a common neural network for non-linear encoding it into the functions' latent space and decoding back to reflectance,without restrictions on the format,size or order of the material data.After an end-to-end training is completed,a common function latent space of these materials can be obtained.Each vector in the space can be represented as a compressed representation for a BRDF function,which can be used to reconstruct reflectance corresponding to a new input through the neural network.In addition,we evaluate the local continuity and global clustering semantics of our function latent space,which makes it possible to interpolate and edit the compressed representation to obtain a new BRDF.In practice,our NPs-based representation can benefit applications in BRDF compression,interpolation,recommendation,and reconstruction.Leveraging this latent space,our encoded BRDFs are highly compact and offer a level of accuracy significantly better than prior methods.