Creating Generative Models from Range Images

Ramamoorthi, 1998

Category: Computer Vision

Overall Rating

3.3/5 (23/35 pts)

While the paper's overall framework relying on hand-crafted hierarchies is largely superseded by data-driven methods, a specific technical idea holds potential: the use of a parameter-space based correspondence and a smooth objective function for fitting structured generative models.
This approach, linking points by their relative position within a parameterized structure rather than geometric proximity, could inform research in learning structured implicit or explicit representations where standard geometric losses are brittle.
It offers a specific, albeit niche, avenue for developing more robust fitting methods for objects well-described by known parameterizations.

This paper's core unconventional potential lies in its approach to geometric model fitting using parameter-space correspondence and smooth, structure-aware objective functions.
A parameter-space based correspondence (the "fractional mapping") where range data points are mapped to model points not by geometric proximity, but by their relative positions within the bounds of the data and the parametric model's parameter space.
Imagine combining this structured, parameter-space approach with modern implicit neural representations.
The robustness of the paper's correspondence method and smooth objective function to noise and missing data (demonstrated on challenging scans) could be invaluable.

The most fundamental issue is the paper's reliance on a model-driven, hand-crafted hierarchy for recognition and representation.
The requirement for a user-defined hierarchy and user-supplied initial guesses for parameter estimation within that hierarchy is a fatal blow to practicality and automation.
The fractional mapping approach, while geometrically intuitive for simple parameterizations, is likely brittle for more complex, non-developable, or non-star-shaped generative models.
Modern researchers should avoid investing significant time into reviving this paper's specific framework because its core assumptions (hand-crafted hierarchies, clean data, manual intervention) are outdated.

Watch