Abstract

There is an open problem in Computer Vision which this talk names the `dolphin problem’. The challenge is to start from multiple photographs of objects from the same class (such as dolphins), where each image sees a different individual in a different pose. This type of collection is typically the result of searching for images that match a certain keyword, for example. The dolphin problem asks whether we can use these images to learn a deformable surface model for objects of the photographed class, with the pose and view parameters for each individual.

This involves an extension of existing surface reconstruction techniques to both multiple poses and multiple views, but the problem is essentially geometric. For example, a formulation of the problem without pose variation is: can we create a fair surface with specified silhouette shapes, where the silhouettes are not necessarily planar cross-sections from the surface?

I will describe an attempt to tackle the dolphin problem using subdivision surfaces as an underlying representation, considering both the challenges and benefits of this strategy. This work was the outcome of an internship at Microsoft Research, Cambridge, in collaboration with Andrew Fitzgibbon. This talk is a rehearsal for a presentation of the same material at a worskhop (`New Trends in Applied Geometry’) the following week.