In this talk, I will address some existing problems in low-level
computer vision and 3D multi-view geometry.
1. On the Morphological Regularizations for Super-resolution image reconstruction:
Why do we care about this problem? Regularization is a well-known technique to convert an ill-posed problem into a well-posed one and obtain a stable solution.
What methods have been studied? Researchers have been studying total variation (TV) based regularization methods for the general inverse problem for last few decades. However, the images reconstructed by these methods sometimes produce ringing artifacts near strong edges.
What do we propose? Based on multi-scale morphology, we develop non-linear regularization methods for edge preserving image reconstruction for general ill-posed inverse problems (in particular Super-Resolution image reconstruction) that can produce better-reconstructed images.
2. On the case of large Hyperedges:
What problem are we trying to solve? Most of the existing works on hypergraph clustering have considered the smallest possible hyperedge size, due to a lack of study into the potential benefits of large hyperedges and effective algorithms to generate them.
What's the answer? Here, we establish that the large hyperedges are better from both theoretical and empirical standpoints.
What’s the approach? We propose a novel guided sampling strategy for large hyperedges, based on the concept of random cluster models. Our method can generate pure large hyperedges that significantly improve grouping accuracy without exponential increases in sampling costs.
What are the implications of the answer? For higher order clustering always practice large hyperedges and use a guided sampling method to sample them.
3. Globally optimal maximum consensus using ASTAR:
Why is this problem so important in 3D vision? Maximum consensus is arguably the criterion of choice for robust estimation tasks in computer vision. Despite its widespread use, optimizing the criterion is still customarily done by randomized sample-and-test techniques.
What's the solution we propose? We aim to change this state of affairs by proposing a very efficient algorithm for global maximization of consensus. Under the framework of LP-type methods, we show how consensus maximization for a wide variety of vision tasks can be posed as a tree search problem.
What's the impact on literature? By returning globally optimal estimates within tens of seconds on typical-sized problems, our approach provides a practical alternative to randomized techniques.
Dr. Pulak Purkait is currently a postdoctoral researcher in University of Adelaide, Australia. He completed his M.Tech and PhD in CS at ISI Kolkata in 2009 and 2013 respectively. His B.Sc. Mathematics and M.Sc. in Applied Mathematics were from the University of Kolkata in the years 2005 and 2007 respectively. His research interests include Image processing, computer vision, 3D multiple-view geometry, machine learning algorithms for scalable 3D-geometric problems.