Cluttered Scene Understanding and Augmentation

L. Li, W.A. Weliamto, and H.S. Seah (Singapore)


Feature Pruning, Clustering, Singular Value Decomposi tion, Collineation, Structure from Simulated Motion, Ac tive Template, Graphical Integration


In this paper, we address the problem of effective 3D clut tered scene understanding and augmentation. These objec tives can be achieved by solving the tasks which can be classified into two parts: clutter handling and real scene augmentation. The basic idea of our methods is to extract planar objects from a real-scene sequence, and then build 3D structure of the planes. Virtual objects can be integrated into natural scenes after recovering planes in the scenes. To achieve this, feature detection and pruning are first used on the target images to reduce clutter in the scene. The pruned scene now contains shape clusters to be matched with a template, which is the known geometry of the target ob ject. Instead of tackling the camera calibration problem, which is a prerequisite process in most vision-based sys tem, we adopt a Structure from Simulated Motion (SFSM) method to get the extrinsic parameters of camera. SFSM, an extension from Structure from Motion (SFM), uses the known geometry of the target object in the image to gen erate a second image. The pose of the camera is then ob tained by using SVD (Singular Value Decomposition) on the collineation, which is the transformation between the target object and second image. The camera motion is re covered from the result, and thus, virtual objects can be placed and aligned coherently with the detected target.

Important Links:

Go Back