Place: Large Lecture Room
Affiliation: Perceiving Systems group at the Max Planck Institute for Intelligent Systems.Tübingen, Germany.
This talk will discuss recent work on object detection. Efficient test time inference is a desirable property of every computer vision algorithm. Especially in structured output spaces tractable inference is an underlying problem that usually. Usually this is tackled by restricting the model class to simpler but efficient methods. In this talk I will present a search based inference technique that allows to perform object detection solely using non-linear kernel functions. The algorithm “learns” the inference problem already during training time, which positively correlates detection speed and classifier accuracy. In other words, the better the detector becomes, the faster it will get. In a second part I will discuss an extension of the well known deformable part model (DPM) to include 3D geometric information. This formulation produces more fine-grained object hypotheses, such as viewpoints and location of individual object parts. The resulting object part predictions are consistent over viewpoint changes.