Physically based rendering follows the physical behaviour of light as closely as
	possible in an effort to predict what the final appearance of a design will be.
	An efficient simulation needs surfaces with accurate physical properties as input data.
	The Bidirectional Reflectance Distribution Function (BRDF) is a fundamental local
	property describing the light reflection on a surface, which exerts an
	influence upon the global illumination of the scene due to interreflections.
														 
	The first section of this dissertation consists in a study of BRDF,
	its physical properties, and a state of the art in BRDF modeling.
	The main conclusion is that even if phenomenological models are common,
	they are not well-suited for a general use. Indeed, they are restricted
	to a specific type of surface and they have a limited angular or spectral range.
	Moreover, simple empirical models cannot handle the behavior
	of complex reflection phenomena. On the other hand, evolved models computing
	the BRDF from explicit surface representations are difficult to implement,
	computationally expensive and their physical parameters often hard to obtain.
	Therefore, a numerical representation based on wavelets is developed
	in this thesis for BRDF measurements modeling because
	measurement is still one of the best way to access BRDF knowledge.
	
	In the second section, the choice of modeling is first discussed and its advantages presented :
	"universal" approach, extensible, speed, data compression, denoising.
	Then, a state of the art of wavelet transforms to BRDF modeling is provided.
	In the order to overcome drawbacks of previous works, a new model
	based on sperated wavelet transforms applied to the directional and
	the spectral component of BRDF is presented.
	It allows a generic representation of the wavelet transform that increases
	compression ability and representation flexibility.
	
	In the third section, the model is tested and evaluated.
	Firstly we perform a non-regressive step in which the new
	model and the previously developed models are compared.
	Then, the modelling error according to the compression rate
	is studied on virtual and real BRDF measurements.
	The model sensitivity to measurement noise is also evaluated, and
	finally the efficiency in term of memory and speed is measured.
	The main conclusion is that the model provides a compact,
	efficient and accurate representation of all kind of BRDF :
	isotropic/anisotropic and monochromatic/spectral BRDF.
	More, the model provides a denoising ability for moderated but realistic noises.
	
	Finally, the model is applied to realistic rendering.
	Thanks to its generic representation of the wavelet transform,
	the model is able to handle other radiometric terms than BRDF :
	spectra, phase functions, and emission distributions.
	Modelling of each term is studied separately in the context of realistic rendering.
	Particularly, an importance sampling scheme is developed in order to
	reduce variance into Monte Carlo based simulations (path tracing, photon mapping).
	At last, we present some optimizations due to the sparse
	wavelet representation for spectral computations.