Grant: Discovery Grants Program - Individual

The goal of computer vision is to develop systems that automatically process visual information. The applications are numerous, ranging from the traditional ones such as industrial inspection and robot navigation, to the newer ones such as video conferencing, and special effects for the movie industry. Markov Random Fields (MRF) and Conditional Random Fields (CRF) are popular probabilistic models for solving challenging labelling problems that are encountered in computer vision. Models that arise require computationally intensive energy minimisation. For many interesting models, the exact minimisation is an NP-hard problem, and only an approximate solution can be found. Thus developing efficient minimisation techniques is essential for obtaining a good solution. I plan to develop more effective methods for minimisation of binary non-submodular energies, which is an NP-hard problem. Binary energies are useful for a variety of problems such as image segmentation, shape priors, etc. Furthermore, any mutli-label energy can be converted to a binary energy. Thus binary non-submodular energies form an important class of energies to handle. In our previous work, we explored two approaches based on the trust region and the auxiliary function frameworks. For most applications, trust region works better than the auxiliary function approach, but auxiliary functions approach is faster. I plan to extend the auxiliary functions approach so that it is as fast but more accurate. I also plan to develop effective optimisation algorithms for densely connected CRFs. Densely connected CRFs are gaining popularity recently, especially since they can be combined with the recently hugely successful deep convolutional neural networks (CNNs) into one system. I plan to develop unified CNN and densely connected CRF models that use efficient minimisation methods and can be trained jointly. Interactive segmentation is a popular tool in medical image processing. Due to uncertainly in the data, and disagreement even among the medical specialists, automatic segmentation is unlikely to overtake user assisted segmentation in popularity. However, it is important to reduce user effort. I plan to develop segmentation tools that require a minimal user assistance in most cases, for example a single click. I plan to continue research on segmentation with shape priors. Shape priors result in more accurate image segmentations since they rule out impossible shape solutions. In previous work I considered rather simple generic shape priors such as convexity, symmetry, etc. I will develop shape priors that are more shape-specific. The proposed research is intended to produce novel optimisation tools and to advance the applicability of optimisation tools for computer vision problems. This, in turn, will lead to an improved performance for the practical problems in computer vision field.