The task of 6D object pose estimation, i.e. of estimating an object position (three degrees of freedom) and orientation (three degrees of freedom) from images is an essential building block of many modern applications, such as robotic grasping, autonomous driving, or augmented reality. Automatic pose estimation systems have to overcome a variety of visual ambiguities, including texture-less objects, clutter, and occlusion. Since many applications demand real time performance the efficient use of computational resources is an additional challenge.
In this thesis, we will take a probabilistic stance on trying to overcome said issues. We build on a highly successful automatic pose estimation framework based on predicting pixel-wise correspondences between the camera coordinate system and the local coordinate system of the object. These dense correspondences are used to generate a pool of hypotheses, which in turn serve as a starting point in a final search procedure. We will present three systems that each use probabilistic modeling and sampling to improve upon different aspects of the framework.
The goal of the first system, System I, is to enable pose tracking, i.e. estimating the pose of an object in a sequence of frames instead of a single image. By including information from previous frames tracking systems can resolve many visual ambiguities and reduce computation time. System I is a particle filter (PF) approach. The PF represents its belief about the pose in each frame by propagating a set of samples through time. Our system uses the process of hypothesis generation from the original framework as part of a proposal distribution that efficiently concentrates samples in the appropriate areas.
In System II, we focus on the problem of evaluating the quality of pose hypotheses. This task plays an essential role in the final search procedure of the original framework. We use a convolutional neural network (CNN) to assess the quality of an hypothesis by comparing rendered and observed images. To train the CNN we view it as part of an energy-based probability distribution in pose space. This probabilistic perspective allows us to train the system under the maximum likelihood paradigm. We use a sampling approach to approximate the required gradients. The resulting system for pose estimation yields superior results in particular for highly occluded objects.
In System III, we take the idea of machine learning a step further. Instead of learning to predict an hypothesis quality measure, to be used in a search procedure, we present a way of learning the search procedure itself. We train a reinforcement learning (RL) agent, termed PoseAgent, to steer the search process and make optimal use of a given computational budget. PoseAgent dynamically decides which hypothesis should be refined next, and which one should ultimately be output as final estimate. Since the search procedure includes discrete non-differentiable choices, training of the system via gradient descent is not easily possible. To solve the problem, we model behavior of PoseAgent as non-deterministic stochastic policy, which is ultimately governed by a CNN. This allows us to use a sampling-based stochastic policy gradient training procedure.
We believe that some of the ideas developed in this thesis,
such as the sampling-driven probabilistically motivated training of a CNN for the comparison of images or the search procedure implemented by PoseAgent have the potential to be applied in fields beyond pose estimation as well.
| Identifer | oai:union.ndltd.org:DRESDEN/oai:qucosa:de:qucosa:31154 |
| Date | 31 August 2018 |
| Creators | Krull, Alexander |
| Contributors | Rother, Carsten, Matas, Jiri, Technische Universität Dresden |
| Source Sets | Hochschulschriftenserver (HSSS) der SLUB Dresden |
| Language | English |
| Detected Language | English |
| Type | doc-type:doctoralThesis, info:eu-repo/semantics/doctoralThesis, doc-type:Text |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0021 seconds