May. 16 2008 10:00AM

May. 16 2008 11:00AM

One of the important tasks of computer vision is automatic 3D scene reconstruction from a given set of images. The fundamental part of this task lies in establishing dense correspondences between the images. The matching should be error free (in terms of mismatches and false-positives) to attain the reconstruction successfully, since these kinds of errors may spoil the reconstruction process completely.

These requirements mean that not the whole of the images should be matched. On the contrary, only those parts of the input images, which are unambiguous, should be interpreted. To achieve that, the matching method must be robust, i.e.\ to be able to reject unreliable data. The quality of the results of these methods (typically) depends on the quality of matching features, since the matching problem is solved directly over them.

Local joint image modelling in stereo matching can bring more discriminable and stable features which reduce the need for algorithms with strong prior models (e.g. continuity). One of the principal quality of the matching features is their shape. In our approach, the matching features are modelled in a 3D space, where they adapt their shape to initial matching hypothesis. Since it cannot be determined directly, we propose a Stratified Dense Matching (SDM) approach. In the experiments we show that our features are able to increase the matching density 3x, matching accuracy 1.8x and occlusion boundary
detection 2x as compared to a fixed-size rectangular windows algorithm.

At the end of our presentation, image processing pipeline that uses SDM will be demonstrated on 3D reconstruction of buildings or cultural artefacts in complex scenes from a set of uncalibrated images.