by Michal Haindl, Jiří Filip and Martin Hatka
Physically correct and realistic visual appearance rendering or analysis of material surface visual properties require complex descriptive models capable of modelling material dependence on variable illumination and viewing conditions. While recent advances in computer hardware and virtual modelling are finally allowing the view and illumination dependencies of natural surface materials to be taken into account, this occurs at the expense of an immense increase in the required number of material sample measurements. The introduction of fast compression, modelling and rendering methods for visual data measurements is therefore inevitable.
The established practice in computer vision, graphics and pattern recognition is to base inferences only on restricted information. Virtual reality applications typically use oversimplified models that cannot even remotely approximate the appearance of real scenes, meaning human observers can easily differentiate between real and virtual scenes. Fortunately, the recent swift development of computer technology has allowed models and tools that seemed theoretical only a decade ago to become feasible as a part of foreseeable future routine processes. Physically correct and accurate material appearance visualization is in high demand. It not only has a huge economic impact in visual safety simulations and virtual design in automotive industry and architecture, but also has large potential in visual scene analytical applications, including health care, security, defect detection and content-based image retrieval. However, many challenging problems still exist, such as efficient measurement of material optical properties, image compression, optimal mathematical representation, unsupervised segmentation and interpretation and many others.
Figure 1: Examples of the realistic rendering of view- and illumination-dependent textures in a car interior, on virtual drapery and on an environmentally lit tablecloth.
We have developed several multidimensional probabilistic models based either on a set of underlying Markov random fields or probabilistic mixtures, which allow physically correct surface lossless representation and modelling, huge measurement space compression (so far unbeaten at up to 1:1 000 000), and even modelling of previously unseen surface data or their editing. These methods are parametric, so they do not require original measurements to be stored. However, such models are nontrivial and suffer from several challenging theoretical problems such as stability, parameter estimation and noniterative synthesis, which must be circumvented.
Alternative approaches using physical reflectance models or sophisticated sampling were also investigated. Regardless of the traits of individual models, they all meet comprehensible requirements such as unlimited seamless surface image enlargement, high visual quality and compression, as well as some less obvious features like strict separation of the analytical and synthesis parts, possible parallelization and implementation in advanced graphics hardware. Unfortunately, there is no ideal universal visual surface mathematical model suitable for all applications or material types. Each of these aforementioned models have their advantages and drawbacks simultaneously, hence optimal measurement as well as modelling depends on both material and intended application, and must be automatically recognized. Surprisingly, the reliable assessment of visual quality is also a difficult task because no usable mathematical criterion exists and such verification requires costly and time-demanding psychophysical visual evaluation. On the other hand, we have successfully applied methods of visual psychophysics to the development of even more efficient material-dependent compression and measurement methods.
We believe that the combination of perceptually optimized measurement and effective mathematical representation of surface appearance is a key to the wide applied usage of realistic view- and illumination-dependent surface material optical measurements.
Links:
http://ro.utia.cas.cz/10cvpr_tutorial/
http://staff.utia.cas.cz/filip/projects/pertex
Please contact:
Michal Haindl - CRCIM (UTIA)
Tel: +420 266052350
E-mail:
