Creating Generative Models from Range Images

Citation

Ramamoorthi, Ravi (1998) Creating Generative Models from Range Images. Master's thesis, California Institute of Technology. doi:10.7907/VZXG-Q639. https://resolver.caltech.edu/CaltechTHESIS:10162018-111851476

Abstract

We describe a new approach for creating concise high-level generative models from one or more approximate range images. Using simple acquisition techniques and a user-defined class of models, our method produces a simple and intuitive object description that is relatively insensitive to noise and is easy to manipulate and edit. The algorithm has two inter-related phases—recognition, which chooses an appropriate model within a given hierarchy, and parameter estimation, which adjusts the model to fit the data. We give a simple method for automatically making tradeoffs between simplicity and accuracy to determine the best model. We also describe general techniques to optimize a specific generative model. In particular, we address the problem of creating a suitable objective function that is sufficiently continuous for use with finite-difference based optimization techniques. Our technique for model recovery and subsequent manipulation and editing is demonstrated on real objects—a spoon, bowl, ladle, and cup—using a simple tree of possible generative models.

We believe that higher-level model representations are extremely important, and their recovery for actual objects is a fertile area of research towards which this thesis is a step. However, our work is preliminary and there are currently several limitations. The user is required to create a model hierarchy (and supply methods to provide an initial guess for model parameters within this hierarchy); the use of a large pre-defined class of models can help alleviate this problem. Further, we have demonstrated our technique on only a simple tree of generative models. While our approach is fairly general, a real system would require a tree that is significantly larger. Our methods work only where the entire object can be accurately represented as a single generative model; future work could use constructive solid geometry operations on simple generative models to represent more complicated shapes. We believe that many of the above limitations can be addressed in future work, allowing us to easily acquire and process three-dimensional shape in a simple, intuitive and efficient manner.

Thesis Files