Here is the abstract of the talk I will be giving at this week’s reading group. Since it is regarding my recent work, I do not have a paper for your perusal.
A Simple Approach to Alignment without Correspondence
This talk will present a new and generalized framework to to align a model with respect to an image. The approach does not explicitly require the nature of imaging or model registration process, illumination and reflectance conditions, surface properties of the scene, determining occlusions and correspondence of features between data from sensors.
The method is based on a new formulation of minimizing a distance measure relating the statistics of two randomly sampled data sets, in our case, a range model and an intensity image. Sensor data have regions best suitable for computing a statistic while parts of it are cluttered or unresolved. The framework exploits the General Crofton Theorem in integral geometry to combine statistics from disparate local regions of the data.
This new approach can accommodate data from different sensors and can be used to align data of same or varied modalities.