Learning Sampling Based Object Pose Estimation: from Analysis-by-Synthesis to Pose Agents and Beyond

Date

May 9, 2017

Time

12:00 PM - 1:00 PM

Speaker

Dipl.-Medieninf. Alexander Krull

Affiliation

Institut für Künstliche Intelligenz

Language

en

Main Topic

Informatik

Other Topics

Informatik

Description

6D object pose estimation is the task of inferring the 3D position and 3D orientation of an object from an input image. This task is important in many applications such as robotics or augmented reality. In this talk, I will discuss a strategy that is applied in many tasks and methods: Sampling a pool of hypotheses, to serve as basis for a search. In the context of 6D object pose estimation we will look at trainable methods for both: sampling such an hypothesis pool and performing the search starting from an hypothesis pool. We will focus in particular on two approaches for learning a search: The first approach ranks hypotheses by learning to compare observed and synthesized images in a probabilistic formulation. It delivers strong improvements in accuracy especially in situations with heavy occlusion. The second approach goes a step further. We learn an algorithm in the form of a reinforcement learning (RL) agent. The RL agent manipulates the hypothesis pool, by iteratively choosing hypotheses for refinement before making a final decision. We believe that RL can play an important role as a tool to learning complex algorithmic pipelines, especially in situations with a limited computational budget. Diese Veranstaltung wird unterstützt von Lehrstuhl Computer Vision.

Last modified: May 9, 2017, 9:42:58 AM