Speaker: Mashhuda Glencross
Title: Creating High Quality Relightable Buildings from Photographs
The human perceptual system is the key to creating effective and believable 3D imagery from photographs. An otherwise accurate model looks "peculiar" if the surface appears to take on the wrong texture or shininess. A significantly less accurate model with the correct surface appearance on the other hand can appear perfectly plausible. In this talk I will illustrate this idea through example ``hallucinated'' 3D models of textured surfaces and show how such models can be transferred to entire building facades.
Although our recovered depth is underestimated, our models contain local depth variations matching the visual and spatial characteristics of the texture over the recovered surface. Based on a study of the plausibility of our results, I will argue that the key to improving acquisition of 3D models from images relies on developing a better understanding of the human visual perceptual system. By exploiting characteristics of this, we can establish both the minimal photographic capture/recovered model fidelity requirements as well as determine how to best minimize the appearance of visual artifacts and errors.
Mashhuda Glencross holds a full faculty post in the Computer Science department at Loughborough University (UK). Her research merges concepts from computer graphics, computer vision, computational photography and visual perception to advance presentation and practical model acquisition processes for game content creation, the built environment and cultural heritage applications. She is an active member of the graphics community and has held a number of committee roles, both technical and organisational, within ACM SIGGRAPH.