Alvey MMI-007 Vehicle Exemplar: Evaluation and Verification of Model Instances

It has proved very difficult to recognise threedimensional objects in natural scenes, based only on two-dimensional features extracted from an image using bottom up methods. One of the major problems is that of relating the two-dimensional information available in the image to three-dimensional entities. This problem can be tackled by attempting to reconstruct knowledge about depth, and hence gain 2V2-D information to work from, this method however is only applicable when fairly precise data, i.e. low-noise images, or multiple image views are available. Alternatively, viewpoint independent features in the image can be identified and used for matching to models. This is simpler but again restricted. However in the approach described here, using a model-based hypothesise-and-lcst strategy, hypothesised two-dimensional instances of threedimensional models can be used to match against directly observable, view-specific image features. This obviates the necessity of obtaining three-dimensional data from the image, and as Lowe points out, viewpoint dependent matching is very much more powerful than any viewpoint invariant method [Lowe 1987J.