Machine und Deep Learning Seminar

Abstract – A Case for Coherent Point Sets as Versatile Object Representation

Recent scene representations combine impressive high fidelity with high efficiency, allowing to reconstruct large-scale scenes in astonishing details. However, they are not suited to model advanced information, such as compositionality or non-rigid objects, and can not be obtained from a small amount of observations. This is in contrast to the requirements of many practical applications in computer vision, such as robotics or 3D modeling, which require to overcome these limitations. This issue is most likely to be solved when using object-level representations.
To this end, this talk will make a case for coherent point sets as efficient and highly effective object representation, finding a good trade-off between flexibility, efficiency and expressiveness. Dr. Jan Eric Lenssen will introduce the general idea and key concepts, before he will present his recent works that apply such point sets in the areas of single-view reconstruction, monocular non-rigid reconstruction and object generation.