Talk by Rudolph Triebel on "Supervised and Unsupervised Learning for Robot Perception"

On Tuesday, August 16th, Rudolph Triebel from the University of Oxford will give a talk about "Supervised and Unsupervised Learning for Robot Perception". The talk will take place in the seminar room on the second floor in Karlstr. 45, Munich.

Abstract:
An important prerequisite for most applications performed with mobile robots is that the robot is able to perceive its environment robustly and reliably. This perception task consists of three major steps: data acquisition, model representation and semantic interpretation. In this talk, I will particularly focus on the semantic interpretation step, in which important information such as object labels or a high-level scene description are automatically added to the environment model. To solve semantic interpretation, a number of different approaches based on supervised learning have been proposed, and I will show some examples from our work at ETH Zurich on pedestrian and car detection using camera and laser data. In addition to that, I will show new approaches, which do not require any hand-labeled training data to extract semantic information. These unsupervised and self-supervised techniques are either based on the detection of repetitive structures in the input data or they rely on information obtained from other sensor modalities such as Inertial Measurement Units (IMU). In the talk, I will show how such non-supervised techniques can be used to discover compound objects such as chairs in 3D range scans or to classify driving scenarios online using vision and IMU data.