Hierarchical Grouping - The Gestalt Assessments Method

Eckart Michaelsen, Michael ArensWorkshops @ ICCV 2017 · DOI

Abstract

This paper presents a mathematical framework for detecting nested hierarchies of symmetry and repetition in images. By applying Gestalt laws—such as proximity and similarity—through continuous assessment functions, the method constructs a "parse-tree" of an image's structure. This approach moves from simple super-pixel primitives to complex, multi-level patterns, effectively capturing the "grammar" of man-made scenes like building facades.

The Challenge of Scene Grammar

While modern computer vision excels at object identification, it often fails to grasp the underlying organizational patterns of a scene. Real-world images, particularly architectural facades, are not merely single mirror images; they contain nested symmetries where reflections are composed of repeating rows (friezes), which are themselves built from smaller symmetric components. Existing methods typically focus on single-level symmetries or require perfectly rectified imagery. This research introduces a flexible, algebraic approach to identify and parameterize these "patterns within patterns" in high-resolution photographs.

Original building facade (Image #34) used for benchmarking hierarchical symmetry detection.

Methodology: The Gestalt-Domain

The proposed method operates in a "Gestalt-domain" where every object is defined by four attributes: location, scale, orientation, and a continuous assessment score (0 to 1). The process follows a bottom-up hierarchy:

  1. Primitives: The image is partitioned into super-pixels using the SLIC algorithm, serving as the base level.
  2. Operations: Two primary operations combine objects: Reflection (|) pairs mirrored objects, and Frieze Repetition ($\Sigma$) groups regularly spaced, similar objects into rows.
  3. Assessment & Search: Continuous mathematical functions calculate scores that are inherited by higher-level structures. A greedy search strategy starts with the highest-scoring pairs and iteratively builds complex patterns.
  4. Clustering: Dominant patterns are identified by clustering reflection axes in the projective plane ($RP^2$).
Hierarchical grouping results showing the detection of friezes (repeating rows of windows) on the building facade.

Key Insights and Results

The framework treats symmetry as a parseable grammar rather than a binary feature. By using continuous assessment functions instead of hard thresholds, the system maintains uncertainty and can rank different structural interpretations.

Qualitative results demonstrate the system's ability to identify level-3 reflection Gestalten—capturing the broad symmetry of a building while simultaneously acknowledging the smaller friezes and window-level symmetries within it. While the method is currently a proof-of-concept with room for optimization in parameter tuning and recognition rates, it successfully generates complex outputs like 4D vector-space clusters and COCO-style symmetry heatmaps.

Conclusion

This hierarchical approach provides a robust tool for structural analysis, moving beyond simple detection to a deeper understanding of scene organization. Future work will focus on refining assessment parameters and integrating additional features like color and texture to enhance performance.

Summary updated

BibTeX
@inproceedings{michaelsen2017hierarchical,
  author = {Eckart Michaelsen and Michael Arens},
  title = {{Hierarchical Grouping - The Gestalt Assessments Method}},
  booktitle = {Workshops @ ICCV},
  year = {2017},
  doi = {10.1109/iccvw.2017.200},
}
Figure