3D Point Correspondence by Minimum Description Length in Feature Space.

Area of Science:

• Computer Vision
• Medical Imaging
• Computational Geometry

Background:

• Point correspondence is crucial for building statistical shape models from 3D surfaces.
• A minimum description length (MDL) objective function balances training errors and generalization.
• Existing MDL methods are effective but do not fully exploit nonlinear properties in correspondences.

Purpose of the Study:

• To adapt the MDL objective function for nonlinear feature spaces to improve 3D point correspondence.
• To develop an efficient optimization method for minimizing the objective function in feature spaces.
• To generalize the MDL objective function to kernel principal component analysis (KPCA) spaces.

Main Methods:

• Adapted the MDL objective function to exploit nonlinear properties in point correspondences using feature spaces.
• Employed Mercer kernels to implicitly define feature spaces.
• Generalized the MDL objective function to KPCA spaces and designed a gradient-descent optimization approach.

Main Results:

• The generalized MDL objective function on KPCA spaces was compared to the original MDL objective function.
• Evaluated performance based on reconstruction errors and specificity.
• Experimental results on 3D human organ shapes demonstrated superior performance of the proposed method.

Conclusions:

• The proposed framework effectively generalizes the MDL objective function to KPCA spaces.
• The new method significantly improves accuracy and generalization in statistical shape modeling.
• This approach offers a powerful tool for analyzing and modeling complex 3D shapes.