Linear Image Features in Stereopsis

Michael KassAAAI 1986

Abstract

This paper introduces a mathematical framework to predict the impact of geometric distortion on stereo matching. By modeling perspective changes as local affine transformations, the author derives confidence intervals that quantify the reliability of image features. This theoretical approach provides a rigorous justification for the success of edge-based stereo algorithms and offers a method for dynamic matching thresholds in 3D reconstruction.

The Challenge of Geometric Distortion

The fundamental task in stereopsis is the correspondence problem: identifying the same physical point across two different viewpoints. However, because cameras are spatially separated, the resulting images undergo geometric distortions such as foreshortening and skewing. A linear filter that produces a specific value in the left image may yield a significantly different value in the right image for the same point, simply due to the change in perspective. This paper addresses how to identify which image features remain stable under these distortions.

Modeling the Transformation

To analyze these effects, the author employs a first-order (affine) approximation of the geometric transformation between views.

The geometric transformation T maps a point p in one image to T(p) in the other, defining the local disparity vector X.

By assuming a statistical distribution of surface orientations, the model establishes limits for horizontal compression and vertical skew. This allows for the derivation of "Geometric Distortion Noise" ($N_{ge}$), an estimate of how much a filter's output is likely to be corrupted by perspective changes. By combining these sensitivity estimates, the algorithm can calculate a local confidence interval for every pixel, identifying unreliable match points before the correspondence process begins.

The Stability of Edges

A significant insight of this work is the mathematical proof regarding the invariance of vertical edges. The analysis shows that for a vertical step edge, the geometric distortion noise essentially vanishes.

Illustration of geometric distortion between two views $I_1$ and $I_2$ for a feature oriented at an angle $\theta$.

This provides a theoretical foundation for why edge-based stereo algorithms, such as the Marr-Poggio-Grimson model, are so effective: edges are naturally more invariant to viewpoint changes than other image features. The paper suggests that matching thresholds should not be global but should instead be adjusted dynamically based on local image structure and its predicted sensitivity to distortion.

Results and Implications

The theory was evaluated using synthetic step edges and real-world aerial stereo imagery of the UBC Acute Care Center. The inclusion of geometric distortion estimates resulted in significantly cleaner depth maps, particularly in distinguishing complex building structures from the ground. While the model relies on an affine assumption that may struggle at sharp depth discontinuities, it represents a major step toward predictive reliability in computer vision, allowing systems to assess their own matching confidence based on local image geometry.

Summary updated

BibTeX
@inproceedings{kass1986linear,
  author = {Michael Kass},
  title = {{Linear Image Features in Stereopsis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure