Event Location: NSH 1305

Abstract: Spatial and temporal alignment are fundamental problems in computer vision that arise naturally in many real-world applications ranging from object localization and visual tracking to image categorization and activity recognition. Most alignment algorithms can be posed as an optimization problem of an energy function over a set of allowable spatial or temporal transformations. Critical to the success of alignment algorithms are the definition of the energy function and the search strategy over the transformation space. Ideally, the former has to enforce that the global minimum of the energy function corresponds to a desirable alignment solution, and the latter has to ensure that the function can be optimized effectively and efficiently. However, these critical factors are often ignored or not fully addressed in the literature; most common approaches to spatial or temporal alignment problems manually design the energy function based on mathematical convenience or ad hoc heuristics, resulting in non-optimal criteria for alignment. In this thesis proposal, we devise alignment algorithms by explicitly addressing the two aforementioned issues in a principled way. We propose margin-based approaches to learn energy functions that: i) achieve their minimum values when spatial objects or temporal events of interest are well aligned; and ii) can be optimized effectively and efficiently. We develop learning formulations for different types of training data: fully supervised, weakly supervised, and unsupervised. In our preliminary work, we show the efficacy of our solution in three scenarios: spatial image alignment of deformable objects, temporal event alignment of facial behavior, and discriminative spatial and temporal alignment of weakly labeled data. Experiments on several real world applications show that our algorithms provide superior results in practice.

Committee:Fernando de la Torre, Chair

Martial Hebert

Carlos Guestrin

Frank Dellaert, Georgia Institute of Technology