Reconstructing representations for real world scenes

Reconstruction of real world scenes and smaller scale objects has received a great research interest in recent years driven by the numerous applications in virtual and augmented reality systems. A system that would provide a general solution to this problem should satisfy several criteria:

• Automatic to limit the reconstruction time and requirement for expert technical knowledge.
• Low-cost to address mass-user applications.
• Accurate to correctly approximate the scene geometry and topology.
• Reliable in the absence of distinct features or surface texture and presence of changes in appearance due to illumination and surface properties.
• Photo-realistic to synthesise a walk-through of an equivalent visual quality to a real video sequence.
• Versatile to reconstruct multiple scale scenes and objects.
• Compact to produce models suitable for transmission and rendering.
• Computationally Efficient to allow simultaneous image capture and model reconstruction.
• Incremental so that the user can identify occluded regions and capture additional images to reconstruct the missing parts of the model.

Unfortunately, no single system currently fulfils all these characteristics and different approaches based on different sensors are currently adopted according to the application priorities. When accuracy is the predominant requirement over cost, compactness and efficiency laser range scanners are the preferred sensor solution. Relative to LRS, reconstruction from photos has also the substantial drawback of being much more sensitive to environmental conditions such as illumination and scene appearance. However, today's digital cameras provide fast and high resolution imaging at affordable rates and so photos can be the favourite input primitive to either image based methods for simple visualisation and navigation applications or geometry based methods for 3D mapping, extensive walkthroughs or video editing.