Psychophysics, Gestalts and Games

José Lezama, Samy Blusseau, Jean-Michel Morel, Gregory Randall, Rafael Grompone von GioiarXiv 2018 · DOI · OpenReview

Abstract

Many psychophysical studies are dedicated to the evaluation of the human gestalt detection on dot or Gabor patterns, and to model its dependence on the pattern and background parameters. Nevertheless, even for these constrained percepts, psychophysics have not yet reached the challenging prediction stage, where human detection would be quantitatively predicted by a (generic) model. On the other hand, Computer Vision has attempted at defining automatic detection thresholds. This chapter sketches a procedure to confront these two methodologies inspired in gestaltism. Using a computational quantitative version of the non-accidentalness principle, we raise the possibility that the psychophysical and the (older) gestaltist setups, both applicable on dot or Gabor patterns, find a useful complement in a Turing test. In our perceptual Turing test, human performance is compared by the scientist to the detection result given by a computer. This confrontation permits to revive the abandoned method of gestaltic games. We sketch the elaboration of such a game, where the subjects of the experiment are confronted to an alignment detection algorithm, and are invited to draw examples that will fool it. We show that in that way a more precise definition of the alignment gestalt and of its computational formulation seems to emerge. Detection algorithms might also be relevant to more classic psychophysical setups, where they can again play the role of a Turing test. To a visual experiment where subjects were invited to detect alignments in Gabor patterns, we associated a single function measuring the alignment detectability in the form of a number of false alarms (NFA). The first results indicate that the values of the NFA, as a function of all simulation parameters, are highly correlated to the human detection. This fact, that we intend to support by further experiments, might end up confirming that human alignment detection is the result of a single mechanism.

Introduction: Bridging Psychophysics and Computer Vision

For decades, the study of vision has been bifurcated between Psychophysics—the qualitative study of human stimulus perception—and Computer Vision—the development of automated detection algorithms. While psychophysics often lacks generic predictive models across diverse scenes, computer vision frequently fails to account for human perceptual nuances such as "masking." Masking occurs when a geometric structure, though mathematically present, becomes invisible to a human observer due to surrounding noise or more dominant patterns.

Figure 1: Visualizing masking by texture and structure.

The authors propose a unified mathematical theory based on Gestalt principles to bridge this gap, aiming for a model that can both implement automatic detection and quantitatively predict human performance.

Methods: The A Contrario Framework

The core methodology relies on the a contrario framework, a formalization of the "Helmholtz Principle." This principle posits that a structure is perceived only when its occurrence is highly unlikely to happen by pure chance.

The Number of False Alarms (NFA)

The "meaningfulness" of a detected structure is quantified by the Number of False Alarms (NFA). A low NFA (typically < 1) indicates a "non-accidental" structure, making it a valid detection. The NFA collapses multiple variables—such as length, density, and precision—into a single scalar metric.

Figure 4: Geometric parameters for NFA calculation.

Gestaltic Games and Masking

To refine the algorithm, the researchers introduced "Gestaltic Games"—an adversarial loop where humans attempt to "fool" the detection algorithm by creating false positives or false negatives. This led to a refined Masking Principle: a gestalt is only considered "masked" if it loses its statistical significance when the elements of a stronger, overlapping gestalt are removed.

Experimental Setup

The study utilized synthetic datasets to maintain parameter control:

  • Dot Patterns: For studying basic alignment and masking.
  • Gabor Stimuli: Over 14,000 stimuli with varying angular jitter and alignment lengths.
  • Online Trials: Data from 7,137 trials collected via a web-based psychophysical experiment.

Results: A Perceptual Turing Test

The results suggest that the NFA serves as an excellent proxy for human detectability. In experiments with Gabor patches, human detection rates followed a sigmoid curve centered at $log_{10}(NFA) = 0$.

Key findings include:

  • Threshold Consistency: Alignments with an NFA < 1 were detected by a majority of subjects, while those with NFA > 1 were generally missed.
  • Cognitive Load: Human response times increased as the NFA approached 1, suggesting that cases difficult for the algorithm are also more mentally taxing for humans.
  • Resolution of Paradoxes: The refined Masking Principle successfully addressed "Kanizsa’s paradox," where sub-structures are hidden within more regular global structures.

Discussion and Future Work

While promising, this framework is currently focused on straight alignments. Future work aims to extend the a contrario model to more complex Gestalt laws, including:

  • Good Continuation: For detecting curved contours.
  • Symmetry and Clustering: Modeling grouping by shape or proximity.
  • Natural Scenes: Moving from synthetic patterns to real-world imagery.

The authors advocate for large-scale online "games" to gather the psychophysical data necessary to further align computational models with human visual perception.

Summary updated

Abstract

Many psychophysical studies are dedicated to the evaluation of the human gestalt detection on dot or Gabor patterns, and to model its dependence on the pattern and background parameters. Nevertheless, even for these constrained percepts, psychophysics have not yet reached the challenging prediction stage, where human detection would be quantitatively predicted by a (generic) model. On the other hand, Computer Vision has attempted at defining automatic detection thresholds. This chapter sketches a procedure to confront these two methodologies inspired in gestaltism. Using a computational quantitative version of the non-accidentalness principle, we raise the possibility that the psychophysical and the (older) gestaltist setups, both applicable on dot or Gabor patterns, find a useful complement in a Turing test. In our perceptual Turing test, human performance is compared by the scientist to the detection result given by a computer. This confrontation permits to revive the abandoned method of gestaltic games. We sketch the elaboration of such a game, where the subjects of the experiment are confronted to an alignment detection algorithm, and are invited to draw examples that will fool it. We show that in that way a more precise definition of the alignment gestalt and of its computational formulation seems to emerge. Detection algorithms might also be relevant to more classic psychophysical setups, where they can again play the role of a Turing test. To a visual experiment where subjects were invited to detect alignments in Gabor patterns, we associated a single function measuring the alignment detectability in the form of a number of false alarms (NFA). The first results indicate that the values of the NFA, as a function of all simulation parameters, are highly correlated to the human detection. This fact, that we intend to support by further experiments , might end up confirming that human alignment detection is the result of a single mechanism.
BibTeX
@misc{lezama2018psychophysics,
  author = {José Lezama and Samy Blusseau and Jean-Michel Morel and Gregory Randall and Rafael Grompone von Gioi},
  title = {{Psychophysics, Gestalts and Games}},
  year = {2018},
  eprint = {1805.10210},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1805.10210},
}
Figure