 |
Kriegman Research Group[ Home ] [ Personnel ] [ Research Projects ] [ Publications ] [ Software ] [ Links ] |
| Curved Object Recognition |
|
Recognition of 3D curved objects using occluding contours in monocular images is a harder problem than its polyhedral counterpart due to the viewpoint dependence of features extracted from the occluding contour.
Therefore, establishing a correspondence between points on the occluding contour to those on the object surface is a daunting task.
As we have shown in [2], [3] we can however establish correspondence between certain points on the occluding contour to curves on the object surface.
We chose to call these HOT (High Order Tangency) curves.
In particular, inflection points of occluding contours are projections of parabolic points on the object surface while bitangent points are projections of limiting bitangent developables.
As we have shown in the paper, by tracking these points through a sequence of images, one can reconstruct the HOT curves.
As an extension of the HOT curve ideas, we can build a representation for 3D Curved Objects (like the duck you see above) which is then used during recognition. Needless to say, in order to perform object recognition in real time, object representation has to efficient and devoid of any redundancies. One way of achieving this is by using invariants. Invariants, are properties of an object that do not change with viewing direction. For out purposes, the invariant properties have to be measured from monocular images. As we all know, an object in 3-D has 6 degrees of freedom with respect to a fixed world coordinate system. These parameters are what dictates the appearance of objects from different viewpoints. However, we could use properties of objects that are independent of all/as many of the above mentioned 6 parameters. We are developing a recognition system that extracts properties (from monocular images) of objects that are invariant to 5 of the 6 rotational and translational degrees of freedom. As a result, objects can be represented as curves in feature space. The invariants we use are parallel-tangent points which are extracted from bitangents and inflection points of occluding contours. Parallel tangent points are points on the occluding contour where the tangent is parallel to a given bitangent or the tangent at an inflection point. During recognition, we extract similar features from the given image and index into our database of models, (i.e., find the model curve that is closest to the measured features in the feature space). For more details on this refer to [4], [5]. Recently, we have been trying to extend the above mentioned Invariant Curve representation to estimate the structure and motion of curved 3D objects from monocular images.
|
| For more details please contact vijay@s3.com. | Updated : Mar 12 2001 |