Incremental model learning and building of unknown objects

Models are very important for many computer vision applications, such as object recognition, detection, and tracking. One can manage to perform the tasks without models, but with models, one can perform the tasks better. These tasks must also handle situation where previously unseen objects appear. In such cases the system must learn the new object or simply derive the descriptions for the object never seen before. In this paper, we propose the creation and use of an online approximate object model using a lidar and vision. The approximate model is represented with the approximate geometry and appearance of 3D objects. We build the model online by tracking the object consistently and estimating the connectivity of cliques representing different views of the same object. To create an model online by the sensor fusion, we investigate two stepwise modeling approaches: 1) motion-based modeling and 2) approximate 3D modeling. Experimental results demonstrate the viability of the proposed time-accumulated 2D and 3D model representation.