Do vision models perceive illusory motion in static images like humans?
Isabella E. Rosario*, Fan L. Cheng*, Zitang Sun, Nikolaus Kriegeskorte · 2026
Read Do vision models perceive illusory motion in static images like humans? by Isabella E. Rosario*, Fan L. Cheng*, Zitang Sun, Nikolaus Kriegeskorte (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-09853.txt.
Introduction
Rosario and three co-authors test optical flow models on Rotating Snakes, testing false motion sight in static images like humans. Most models fail to make flow fields matching human sight. Under mock saccadic eye shifts, only the human-like Dual-Channel model shows wanted spinning motion. Brightness-based and color-feature-based signals feed here, and a looping focus tool is key for joining cues. The results highlight a mismatch between human visual motion work and current optical-flow models, giving insights for future systems matching human sight better.
Catalog record
| CREATOR | Isabella E. Rosario*, Fan L. Cheng*, Zitang Sun, Nikolaus Kriegeskorte |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66002 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.09853
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Do vision models perceive illusory motion in static images like humans? Isabella E. Rosario*, Fan L. Cheng*, Zitang Sun, Nikolaus Kriegeskorte Abstract Understanding human motion processing is essential for building reliable, human-centered computer vision systems. Although deep neural networks (DNNs) achieve strong performance in optical flow estimation, they remain less robust than humans and rely on fundamentally different computational strategies. Visual motion illusions provide a powerful probe into these mechanisms, revealing how human and machine vision align or diverge. While recent DNN-based motion models can reproduce dynamic illusions such as reverse-phi, it remains unclear whether they can perceive illusory motion in static images, exemplified by the Rotating Snakes illusion. We evaluate several representative optical flow models on Rotating Snakes and show that most fail to generate flow fields consistent with human perception. Under simulated conditions mimicking saccadic eye movements, only the human-inspired Dual-Channel model exhibits the expected rotational motion, with the closest correspondence emerging during the saccade simulation. Ablation analyses further reveal that both luminance-based and higher-order color–feature–based motion signals contribute to this behavior and that a recurrent attention mechanism is critical for integrating local cues. Our results highlight a substantial gap between current optical-flow models and human visual motion processing, and offer insights for developing future motion-estimation systems with improved correspondence to human perception and human-centric AI. 1 1 footnotetext: These authors contributed equally to this work. 2 2 footnotetext: Corresponding author: [email protected]. 1 Introduction Interpretation of visual motion information is crucial for successfully navigating and interacting with complex, dynamic environments, including those encountered in autonomous driving, object tracking, robotics, and video interpolation. Humans possess a sophisticated visual system capable of performing diverse tasks with remarkable stability and robust generalization to new environments, while consuming little energy. Although state-of-the-art computer vision models now surpass human performance on many motion estimation benchmarks Yang et al. 2023 , they exhibit well-documented limitations. For example, deep neural networks (DNNs) for optical flow often struggle with occlusions, large displacements, and non-rigid motion, and can behave unpredictably under experimental conditions that humans handle robustly Ren et al. 2017 ; Dosovitskiy et al. 2015 ; Chen and Poullis 2022 ; Ranjan et al. 2019 . These discrepancies suggest that the human visual system relies on computational strategies not captured by models that track per-pixel displacements alone. A deeper understanding of the principles underlying human motion perception may therefore inspire more robust and efficient machine vision systems. One promising direction is to evaluate models through the lens of human perception. Illusory motion phenomena provide a powerful testbed for adjudicating among computational models Weiss et al. 2002 ; Watanabe et al. 2018 ; Kobayashi et al. 2022 ; Kirubeswaran and Storrs 2023 . From both engineering and scientific perspectives, models that reproduce human perceptual biases are valuable: they may generalize better in real-world settings and offer mechanistic insight into the computations of the visual system Yang et al. 2025 . Recent neuroscience-informed motion models incorporating motion-energy filtering, higher-order pathways, and recurrent integration Sun et al. 2023 ; Tangemann et al. 2024 ; Sun et al. 2025 have successfully replicated classic motion illusions such as reverse-phi and the barber-pole effect, suggesting potential shared computational principles with biological vision. However, it remains unclear whether modern DNNs can reproduce anomalous-motion illusions such as the Rotating Snakes Kitaoka and Ashida 2003 ; Otero-Millan et al. 2012 , whose perceptual dynamics depend on subtle luminance asymmetries and fixational eye movements Lotter et al. 2016 ; Watanabe et al. 2018 ; Kobayashi et al. 2022 ; Kirubeswaran and Storrs 2023 . This work addresses this gap using an in silico psychophysical approach to systematically probe the correspondence between deep models and human motion perception. Our contributions are summarized below: • We comprehensively evaluate a broad set of representative DNN-based and human-inspired motion models to assess their ability to reproduce illusory motion in static images , with a primary focus on the Rotating Snakes . • We propose a unified experimental pipeline and validation benchmark that integrates controlled comparisons, simulated viewing conditions, and quantitative evaluation metrics. Under specific conditions, some neuroscience-inspired models exhibit human-like motion percepts that are not reproduced by conventional DNNs. • We perform systematic probing, control and ablation analyses to identify the computational components responsible for illusion replication. Our results highlight key mechanisms, such as dual-channel motion processing, oculomotor transients, and recurrent integration, may support more human-like, robust generalization in complex visual environments. 2 Experimental Design and Method We analyzed anomalous motion illusions (i.e., illusory motion in static images), focusing on the Rotating Snakes illusion Kitaoka and Ashida 2003 . 2.1 Image Design 2.1.1 Rotating Snakes Illusion The Rotating Snakes variant is among the most extensively characterized anomalous motion illusions in human psychophysics and primate neurophysiology Kitaoka and Ashida 2003 ; Conway et al. 2005 ; Atala-Gérard and Bach 2017 ; Bach and Atala-Gérard 2020 ; Fermüller et al. 2010 ; Kitaoka 2014 ; Murakami et al. 2006 ; Mather and Cavanagh 2025 ; Otero-Millan et al. 2012 ; Uesaki et al. 2024 ; Ashida et al. 2012 . It induces highly predictable direction of perceived rotation for human observers, and its parametric stimulus construction affords precise experimental control, allowing rigorously matched illusion and control sets and clean parametric sweeps. We generated Rotating Snakes illusion and control images (Figure 1 ) following established stimulus construction procedures Uesaki et al. 2024 . Each illusory image consisted of six concentric rings, each containing 24 repeating elements. Images were rendered using either grayscale (G), blue–yellow (B-Y), or red–green (R-G) color schemes, all of which have been shown to reliably induce counterclockwise rotation for most human observers. Control images were created by permuting the luminance or color sequence within each repeating unit, thereby eliminating the local luminance gradients responsible for driving the illusion Atala-Gérard and Bach 2017 . These controls preserved the overall spatial layout and chromatic composition while disrupting the specific asymmetric micro-patterns known to generate perceived rotation. Figure 1 : Rotating Snakes illusory and control images. Top row: illusory images in grayscale (G), blue–yellow (B-Y), and red–green (R-G) color schemes, each producing a robust perception of counterclockwise rotation in human observers. Bottom row: corresponding control images created by permuting the luminance or color sequence within each repeating unit, which fails to induce perception of motion while preserving global layout. 2.1.2 Other Anomalous Motion illusions Beyond the Rotating Snakes illusion, we also analyzed a broader set of anomalous motion illusions. These patterns are grouped according to Kitaoka’s tentative classification of anomalous motion illusions Kitaoka 2025 , which organizes them by the image features that drive illusory motion. Figure 7 show the specific exemplars used in our experiments. Peripheral drift illusions correspond to a broader family that includes the Rotating Snakes illusions Faubert and Herbert 1999 . Observers report that the rings rotate or “creep” around the center when they move their eyes, blink, or experience a sudden stimulus onset. The exemplars sample several geometries ( e.g . , nested circular arrangements, radial spokes, and logarithmic spirals) but share the same core design principle: a repeating sequence of unequal luminance steps that biases the motion system under transient retinal slips. In central drift illusions Kitaoka and Ashida 2004 , radial-shaped sectors appear to rotate even