Do Vision Models Perceive Illusory Motion in Static Images Like Humans?¶

Conference: CVPR 2026 arXiv: 2604.09853 Code: Available Area: Visual Perception / Computational Neuroscience Keywords: motion illusion, optical flow models, human vision, rotating snakes illusion, biologically-inspired models

TL;DR¶

This paper systematically evaluates a range of optical flow models on static-image motion illusions such as the Rotating Snakes, finding that only the biologically-inspired Dual-Channel model reproduces the human-perceived rotational motion under simulated saccade conditions.

Background & Motivation¶

Background: DNNs have surpassed human performance on optical flow benchmarks, yet robustness gaps remain. Visual motion illusions provide a powerful tool for probing human–machine differences, but existing studies have focused primarily on dynamic illusions (e.g., reverse-phi), leaving static-image illusions underexplored.

Limitations of Prior Work: The Rotating Snakes illusion—in which humans strongly perceive rotational motion in a completely static image—has not been assessed in terms of whether existing optical flow models can reproduce it. The illusion depends on subtle luminance asymmetries and fixational eye movements.

Key Challenge: Standard DNN optical flow models achieve strong benchmark performance, yet it remains unclear whether their computational strategies share fundamental principles with the human visual system.

Goal: To evaluate the ability of representative DNN and biologically-inspired motion models to reproduce static-image motion illusions, and to identify the key computational components responsible.

Key Insight: An in silico psychophysics approach is adopted, systematically comparing 10 motion estimation models within a unified experimental pipeline.

Core Idea: Dual-channel motion processing, transient signals from eye movements, and recurrent integration are the critical mechanisms for reproducing human-like motion perception.

Method¶

Overall Architecture¶

(1) Generate Rotating Snakes illusion images and control images (three color schemes: grayscale / blue–yellow / red–green); (2) evaluate 10 models under both static and simulated saccade conditions; (3) conduct ablation analyses to identify key components.

Key Designs¶

Unified Experimental Pipeline:
- Function: Enable fair comparison across architectures under controlled conditions.
- Mechanism: All models use official pretrained weights and are evaluated on identical illusion/control images. Simulated saccades are produced by translating images to generate transient retinal slip.
- Design Motivation: Ensure that observed differences are attributable to model architecture rather than training or evaluation discrepancies.
Simulated Saccade Condition:
- Function: Replicate the physiological conditions under which humans view the Rotating Snakes.
- Mechanism: Human perception of the illusion requires transient signals provided by fixational eye movements such as saccades. Image translation is used to simulate this retinal slip.
- Design Motivation: Psychophysical studies show the illusion is substantially attenuated under fixed gaze; eye movements are the key trigger.
Ablation Analysis:
- Function: Identify the computational components essential for reproducing the illusion.
- Mechanism: Systematic ablations of the Dual-Channel model examine: (1) the contribution of luminance-based motion signals; (2) the contribution of higher-order color–feature motion signals; (3) the role of the recurrent attention mechanism.
- Design Motivation: Determine which computational principles are necessary for human-like motion perception.

Loss & Training¶

This work is purely inference-based; no training is involved. All models are used with their original pretrained weights.

Key Experimental Results¶

Main Results¶

Model Type	Static Condition	Saccade Condition	Reproduces Illusion
Multi-scale DNN (FlowNet, etc.)	No rotational flow	No rotational flow	✗
Recurrent-decoder DNN (RAFT, etc.)	No rotational flow	No rotational flow	✗
Dual-Channel (biologically-inspired)	Weak signal	Expected rotational motion	✓

Ablation Study¶

Configuration	Key Metric	Notes
w/o luminance channel	Illusion weakened	Luminance signals contribute significantly
w/o color–feature channel	Illusion weakened	Higher-order signals also contribute
w/o recurrent attention	Illusion disappears	Critical for integrating local cues
Full Dual-Channel	Strongest agreement	All components act synergistically

Key Findings¶

The majority of DNN optical flow models entirely fail to produce human-consistent motion flow fields on static images.
The Dual-Channel model exhibits the expected rotational motion only under simulated saccade conditions; the effect is also weak under the static condition.
The recurrent attention mechanism is the critical component for integrating local cues into global rotational percepts.

Highlights & Insights¶

Motion illusions as model diagnostic tools: Leveraging human perceptual biases to distinguish models that "work" from those that "work like humans."
Validation of biologically-inspired computational principles: Dual-channel motion processing, eye-movement transients, and recurrent integration constitute three transferable design principles.
Implications for robust visual system design: Models capable of reproducing human perceptual biases may also exhibit greater robustness in real-world settings.

Limitations & Future Work¶

Only a limited variety of motion illusion types are tested.
The optical flow estimation performance of the Dual-Channel model is not benchmarked against mainstream DNNs.
Only zero-shot inference is analyzed; whether fine-tuning could enable DNNs to learn to reproduce the illusion remains unexplored.

vs. standard optical flow benchmarks: Strong benchmark performance does not imply alignment with human vision; motion illusions provide a complementary evaluation dimension.
vs. reverse-phi studies: Reverse-phi is a dynamic illusion, whereas Rotating Snakes is a static illusion, placing higher demands on the model.

Rating¶

Novelty: ⭐⭐⭐⭐ First systematic evaluation of static motion illusions in computational vision.
Experimental Thoroughness: ⭐⭐⭐⭐ 10 models × multiple conditions × ablation analyses.
Writing Quality: ⭐⭐⭐⭐ Interdisciplinary research well organized.
Value: ⭐⭐⭐ Informative for optical flow model design, though practical applications are limited.