A linear method to derive 3D projective invariants from 4 uncalibrated images

Area of Science:

• Computer Vision
• Computational Geometry
• 3D Reconstruction

Background:

• Existing methods for computing 3D projective invariants, like Quan's method, are nonlinear and complex, often yielding multiple solutions.
• Previous research suggested linear solutions using six points in five images, but this paper explores a more efficient approach.

Purpose of the Study:

• To develop a direct and linear method for computing projective invariants of 3D points from four uncalibrated images.
• To address the complexity and ambiguity issues associated with traditional projective invariant computation methods.

Main Methods:

• Representing two 3D points using a basis of four other points within the set.
• Deriving a system of four bilinear equations for three unknown projective invariants through linear transformations.
• Solving the system of equations linearly and uniquely.

Main Results:

• A novel method is presented to compute projective invariants of 3D points directly from four uncalibrated images.
• The method successfully solves a system of bilinear equations linearly and uniquely, overcoming the complexity of nonlinear systems.
• Explicit formulas for the solutions are provided, simplifying the projective reconstruction process.

Conclusions:

• The proposed method offers a significant advancement in computing 3D projective invariants, providing a direct, linear, and unique solution.
• The findings suggest that six points and four images represent a natural configuration for projective reconstruction.
• This research has implications for various computer vision applications requiring accurate 3D scene understanding from uncalibrated imagery.