You are here
Visual Scene Analysis
The visual scene analysis group conducts research in the field of mathematical image processing, computer vision, and machine learning with a special focus on the development of energy minimization methods for improving and analysing digital images. In this context we are interested in convex relaxations, functional lifting and multiscale methods as well as the development of efficient convex and nonconvex optimization algorithm for the solution of the corresponding energy minimization problems. Please visit our publications page to find out more about our recent research.
In the summer semester 2017 we will teach a master course on convex optimization for computer vision. Optimization is a quite exciting field. Having a thorough understanding of the algorithms that work for convex functions provides a good basis even for nonconvex optimization. Even more importantly, optimization is needed everywhere - not just in computer vision!
Feel free to contact me at email@example.com if you are interested in writing your Bachelor or Master thesis in the field of image processing or computer vision! We have several exciting topics! Please find a list of examples below. I am happy to explain these (sometimes technically sounding) topics in more detail.
- In deep learning, e.g. familiarize yourself with a deep learning framework such as https://www.tensorflow.org, or http://deeplearning.net/software/theano, or https://caffe2.ai/ and improve certain image reconstruction tasks with semantic information by following some idea of https://arxiv.org/pdf/1701.01698.pdf.
- In 3D reconstruction, e.g. familiarize yourself with the relation of images and the camera pose to real world coordinates and try to reimplement an energy minimization approach to surface reconstruction similar to https://www.youtube.com/watch?v=TGg-ujjSsOM (but not in real-time). Extensions of such techniques are able to map large scale geometric features, see e.g. https://www.youtube.com/watch?v=GnuQzP3gty4
In optimization, e.g.
- on joint optimization methods for biconvex or nonconvex problems such as blind deblurring or super resolution. Some suprising results show that there is a gap between good results and faithful optimization (http://www.cvg.unibe.ch/dperrone/tvdb/), which could be investigated in more detail!
- on optimizing deeply nested functions (as frequently arising in deep learning applications) with proximal splitting methods.
- on efficient ways to optimize a certain class of constrained convex functions whose proximal operator does not have a closed form.
- on optimizing nonconvex functions by representing the underlying problem in a higher dimensional space, also known as lifting. See https://arxiv.org/pdf/1512.01383.pdf for an example of a recent work on this topic. I can give clear instructions of how a project on a related topic can look like.
- In combining optimization and deep learning methods, e.g. by studying and extending our recent work on using denoising networks as proximal operators, see https://arxiv.org/abs/1704.03488.
- On multiscale methods for inverse problems, extending the current theory on nonlinear spectral decompositions (see, e.g. https://arxiv.org/pdf/1510.01077.pdf). Such a project could be application driven (https://arxiv.org/pdf/1703.08001.pdf), or based on some mathematical analysis of generalized eigenfunctions, see https://arxiv.org/pdf/1601.02912v1.pdf
If you have ideas for your own research project, let me know!