Visual object recognition is a highly complex information processing task that humans seem to perform effortlessly. Understanding visual object recognition is an important scientific problem, and visual object recognition by computer has many applications in areas such as robotics, navigation, human-computer interaction, handwriting recognition, and document analysis. Particularly remarkable about humans is their ability to learn and adapt to visual input, often from unlabeled data.

In my work, I investigate visual object recognition by building computational and statistical models and testing them on real data. In the first part of my talk, I will describe a set of computational tools and algorithms based on branch-and-bound methods that allow geometric matching and optimization problems occurring in visual object recognition to be solved efficiently and robustly. In the second part, I will describe statistical models of the variability of object appearance, as well as methods for learning such models from labeled and unlabeled data. I will provide illustrations of the technique with applications to trainable 3D object recognition, handwriting recognition and OCR, and document image analysis.