Brody Kutt¹, Zygmunt Pizlo¹; ¹University of California, Irvine

How humans form structured perceptions of visual stimuli and judge objectness are long-studied questions in psychophysics. According to the Gestalt laws of perception, the visual system ignores low-level details. It extracts global, simplified symbolic knowledge, which is critical to higher-level cognition, such as concept formation and reasoning. To predict what humans are likely to perceive, prior work cast the problem as a search task, often using complex, hand-crafted scoring procedures that were limited to line drawings. In contrast, this work introduces a closed-form algorithm founded on domain-general visual information processing operations without requiring optimization, search, or training data. MERSymm is an algorithm for assessing the likelihood of a percept grounded in the idea that human judgments of objectness reflect compressible regularities. MERSymm takes as input a mathematical structure, which is a candidate percept of an image. MERSymm detects Euclidean symmetries as indicators of information redundancy using a parallelizable multiscale pyramid and produces a prior probability of the percept for Bayesian inference. We evaluate MERSymm and off-the-shelf compression algorithms on a custom benchmark. Algorithms must identify which perceptual object in a pair exhibits properties associated with objectness (organized into eight Gestalt principles: symmetry, closure, proximity, similarity, continuity, uniform connectedness, common region, and figure-ground). MERSymm detects these eight Gestalt principles nearly perfectly in their basic presentations (currently 99.28% ± 0.011% detection accuracy), unifying them for the first time as arising from a symmetry-based procedure. MERSymm also remains robust under different noise conditions. It captures global compressible regularities that existing compression algorithms miss. Existing algorithms achieve 56%-74% average accuracy. MERSymm is an open-domain, massively parallelizable, and generally applicable algorithm for mirroring human judgements of objectness based only on fundamental visual processing operations. This work has implications for Bayesian models of perception and developing human-like, neurosymbolic AI and computer vision systems.