Planned intervention: On Wednesday April 3rd 05:30 UTC Zenodo will be unavailable for up to 2-10 minutes to perform a storage cluster upgrade.
Published September 25, 2013 | Version v1
Poster Open

A generic model architecture for perceptual grouping

  • 1. Laboratories of Biological and Experimental Psychology, KU Leuven (Belgium)
  • 2. Laboratory of Experimental Psychology, KU Leuven (Belgium)
  • 3. Laboratory of Biological Psychology, KU Leuven (Belgium)

Description

Gestalt grouping principles provide important cues for organizing visual inputs into coherent percepts. While a number of models for each grouping principle have been proposed, a unifying architecture that would flexibly incorporate all of them is still lacking. Moreover, many of these models operate in mathematical terms, rendering them unsuitable for a wider testing on real input images. Here we present a generic biologically plausible unsupervised architecture, which we term gmin (https://github.com/qbilius/gmin), geared towards a unified implementation of the grouping principles. The model operates by extracting a certain feature of interest from an input image (e.g., edges) and selecting maximally informative units to decrease computational load; then it proceeds to iteratively compute grouping strength between all pairs of units and update informative unit selection using a particular grouping principle (e.g., good continuation). This process is repeated for all features of interest. Finally, a segmentation of an image into (provisional) objects is provided by an unsupervised learning procedure on the computed grouping strengths, such as clustering or belief propagation. We illustrate the validity and the general usefulness of our approach by testing it on a few very different grouping displays. Using the cues of proximity, good continuation, or orientation similarity, we were able to obtain a satisfying model performance on the classic Gestalt displays of dot lattices, contour integration, and texture segregation.

Files

Kubilius_et_al_gmin_Bernstein2013.pdf

Files (3.1 MB)

Name Size Download all
md5:fcc16649b67528a4c0375aca95a6b5a7
3.1 MB Preview Download