Introduction

In this section, we consider view dependent mapping as an effective method for rendering a scene that has been reconstructed from photographs. This form of texture mapping is really effective when the model closely conforms to the actual structure of the scene.
Let us first consider how a single image is projected onto the model and then consider multiple images being projected onto it.

Projecting a single image

The process of texture-mapping a single image can be viewed as replacing each camera with a single projector that projects the original image onto the model. When the model is not convex, some parts of the model are going to shadow the others. Such shadowed regions could be determined with any object space surface visibility algorithm.

Compositing multiple images

In general, each photograph only views a portion of the model. So, multiple images are required to render a particular model. Some pixels are visible in one view whereas some pixels are visible in more than one view. If a pixel is mapped in only one rendering, then that value is taken otherwise the composite value is taken.

The problem occurs if the projected images do not agree where they overlap. The images will not agree if there is some unmodelled geometric detail in the building. The best image to view is one that is closest to the rendered view. However, always using the image closest in angle at every pixel means that the neighbouring rendered pixels may be sampled from different images. This gives rise to seams in the rendering. To avoid this problem, averaging is carried out.

Even with averaging, neighbouring pixels can still appear in different views at the boundary of two regions as the contribution of an image outside the region is zero. This could be overcome by having the pixel values ramped down near the pixel boundary. This eliminates most of the unwanted seams in the rendering.

By considering regions over the image rather than individual pixels, a real time version of view dependent texture mapping can be achieved.

Back to the main screen


Copyright - Sudhir R Kaushik (sudhir@cs.wpi.edu)