An Introduction to Recognition Using Surfaces

The surface is the boundary between object and non-object and is the usual source and limit of perception. As such, it is the feature that unifies most significant forms of non-invasive sensing, including the optical, sonar, radar and tactile modalities in both active and passive forms. The presence of the surface (including its location) is the primary fact. Perceived intensity is secondary - it informs on the appearance of the surface as seen by the viewer and is affected by the illumination and the composition of the surface.

Grimson, in his celebrated book ``From Images to Surfaces: A Computational Study of the Human Early Visual System'' [77], described an elegant approach to constructing a surface representation of a scene, starting from a stereo pair of intensity images. The approach, based substantially on Marr's ideas [112], triangulated paired image features (e.g. edge fragments) to produce a sparse depth image, and then reconstructed a complete surface, giving a dense depth image. Though this is not the only way to acquire such a surface description, (e.g. laser ranging and structured light are practical alternatives), there have been only a few attempts at exploiting this rich data for object recognition.

Previous research in object recognition has developed theories for recognizing simple objects completely, or complex objects incompletely. With the knowledge of the ``visible'' surfaces of the scene, the complete identification and location of more complex objects can be inferred and verified, which is the topic of this book. Starting from a full surface representation, we will look at new approaches to making the transformation from surfaces to objects. The main results, as implemented in the IMAGINE I program, are:

• Surface information directly provides three dimensional cues for surface grouping, leading to a volumetric description of the objects in the scene.
• Structural properties can be directly estimated from the three dimensional data, rather than from two dimensional projections.
• These properties plus the generic and structural relationships in the model base can be used to directly invoke models to explain the data.
• Using surfaces as both the model and data primitive allows direct prediction of visibility relationships, surface matching, prediction and analysis of occlusion and verification of identity.
• Moderately complex non-rigidly connected structures can be thoroughly recognized, spatially located and verified.