PhD Defense by Vladimir Haltakov

Speaker: Vladimir Haltakov Date: Friday, September 7, 2018 Time: 09:00 Location: 01.07.014

Learning Context For Semantic Segmentation And Applications

Abstract:

Nowadays, cameras are an integral part of many devices and systems - from mobile phones to autonomous vehicles and from medical robots to surveillance cameras. While an active field of research, the task of understanding a camera image still poses significant challenges and is not solved in general. A key in interpreting camera images correctly is not to focus on individual image areas or objects, but to use context from the whole image in order to resolve ambiguities - something that we humans are very good at.

This thesis is built around the idea of learning context - we present several novel methods that extend the traditional semantic segmentation algorithms in order to allow them to incorporate context relations. We present a new classifier-based pairwise potential that can be integrated in a CRF segmentation framework, extending its abilities to learn local context relations. For learning both local and global context, we presented another method, relying on a classifier chain and novel feature formulation based on geodesic neighborhoods. We also show how this model can naturally handle both texture and 3D information and can also run in real-time.

Another approach for providing semantic segmentation methods with additional information is synthetic ground truth data. We present a generic framework for generating high-quality ground truth data, including camera images, pixelwise semantic labels, depth and optical flow data, based on a driving simulator.

Finally, we show how the proposed methods can be used in two real world automotive applications, in which semantic segmentation and a good context understanding are very important - parking space detection and traffic lights detection.