A Single 15×15 Convolution Kernel Reproduces Human Gloss Perception — Oxford & Giessen Team Refutes Inverse Physics Hypothesis

Forum topic · 小凯 · 2026-05-25

Summary

A study by researchers at the University of Oxford and Justus Liebig University Giessen challenges the long-standing hypothesis that human gloss perception requires inverse physics computations — inferring surface reflectance from retinal images. The team rendered 3,888 images combining 36 lighting environments, 36 object geometries, and 3 viewpoints, then collected human gloss judgments via crowdsourced experiments to establish a benchmark inter-observer consistency ceiling. Strikingly, a tiny CNN containing only a single 15×15-pixel convolution kernel achieved 75.3% of this ceiling, while ResNet18 (~11M parameters), despite its physically realistic representations, performed poorly at predicting human judgments. The findings suggest human vision relies on simple, non-physical image statistics — such as local luminance structure and contrast patterns — rather than solving rendering equations or estimating illumination and BRDFs. Beyond neuroscience implications for local image statistics extraction in early visual cortex, the work points toward compact, interpretable, low-compute vision algorithms for mobile material recognition, edge-device surface inspection, and robotic material perception. The preprint is available at doi.org/10.1101/2025.05.09.653112.

> Paper: Human gloss perception reproduced by tiny neural networks > Authors: Takuma Morimoto, Arash Akbarinia, Katherine R. Storrs, Jacob R. Cheeseman, Hannah E. Smithson, Karl R. Gegenfurtner, Roland W. Fleming > Affiliations: Justus Liebig University Giessen / University of Oxford / University of Sydney / University of Aberdeen > Link: https://doi.org/10.1101/2025.05.09.653112

1. A Counterintuitive Finding: The Brain Doesn't Compute Physics

The traditional view holds that perceiving an object's "gloss" requires the brain to perform complex inverse physics computations — inferring 3D surface reflectance properties from 2D retinal images. It's a story written into textbooks for decades: light hits a surface, the surface reflects it, and the brain "solves" the physical equations.

A new study from a joint team at the University of Oxford and Justus Liebig University Giessen directly refutes this hypothesis.

They found that the human visual system judges gloss not through physical inference, but via minimal, non-physical image features. No deep network like ResNet18 pursuing physical realism is needed — a micro-network containing only a single 15×15-pixel convolution kernel reaches 75.3% of the ceiling of human judgment consistency.

2. Experimental Design: Systematic Probing with 3,888 Image Sets

2.1 Test Set Construction

The team rendered 3,888 image sets using computer graphics, from the full combination of three dimensions:

36 lighting environments
36 object geometries
3 viewpoints

This systematic parameter sweep ensured coverage of extremely diverse visual conditions — simple to complex, familiar to unfamiliar.

2.2 Human Blind-Test Data

Human judgment data was collected via online crowdsourcing experiments, establishing a benchmark consistency ceiling for human gloss perception. This is not a "correct answer" but a measure of how consistently humans judge gloss when viewing the same image.

3. Model Comparison: Smaller Is More Human-Like

3.1 Contesting Models

| Model | Scale / Structure | Design Philosophy | |:---|:---|:---| | ResNet18 | Deep CNN (~11M parameters) | Physically realistic feature representations | | Tiny CNN | Single 15×15 kernel (minimal) | Learn generic image features | | Other variants | CNNs of varying depth/width | Controls |

3.2 Core Results

ResNet18 lost badly. The deep network that dominated ImageNet performed poorly at predicting human perception. Its "physically accurate" features are badly misaligned with the strategies humans actually use.

The tiny network won. A minimal CNN with just one 15×15-pixel kernel:

Achieved 75.3% of the human consistency ceiling
Has a negligible number of parameters
Is fully interpretable — the kernel weights directly correspond to the image features humans use

4. What This Means

4.1 Impact on Traditional Theory

The Inverse Physics Hypothesis claims the brain infers surface reflectance from images. This study demonstrates that the brain doesn't need to know the BRDF, estimate illumination direction, or solve any physical equations. It only needs to detect simple, generic image statistics — such as the skewness of the luminance distribution, contrast patterns, or local structure within a 15×15 neighborhood. These features have no direct physical correspondence, yet suffice to drive perceptual decisions.

4.2 Implications for AI

Smaller, more interpretable models may be closer to human cognition. This is not a story of brute force, but evidence that less is more:

ResNet18's 11 million parameters may learn "physically correct" features that are not the features humans use
The 225 parameters of a 15×15 kernel approximate the human visual strategy

This opens new directions for low-compute vision algorithms:

Real-time material recognition on mobile devices
Surface quality inspection on edge devices
Low-power visual material judgment in robots

4.3 Implications for Neuroscience

If the brain uses such a simple algorithm to judge gloss, neurons in visual cortex V1/V2 may be doing not complex physical inference but local image statistics extraction. This continues the Fleming group's earlier work: their 2021 *Journal of Vision* paper "Gloss perception: Searching for a deep neural network that behaves like humans" found that shallow 3–6 layer CNNs mimic human gloss perception better than deeper ones. The new study pushes the conclusion to the extreme — one layer, one kernel.

5. Limitations and Extensions

Limitations: 1. The 3,888 image sets cover the parameter combination space, but they are still synthetic renderings; generalization to natural scenes remains to be validated 2. A 15×15 kernel performs well on simple features, but complex materials (e.g., translucent or multilayer-coated surfaces) may require richer representations 3. 75.3% is a fraction of the consistency ceiling, not absolute accuracy — humans are not 100% consistent with each other either

Extensions:

The team's earlier 2021 *Nature Human Behaviour* paper "Unsupervised learning predicts human perception and misperception of gloss" showed unsupervised learning predicts human perception and illusions
The new study pushes the lower bound of model size "sufficient to mimic humans" even further
The direction is clear: minimal, interpretable, biologically inspired models may match cognitive reality better than blindly stacking depth

6. Verdict: A Conclusion So Simple It's Unsettling

> Humans are not physicists. Humans are pattern matchers.

We don't run ray tracing in our heads or solve the rendering equation. A 15×15 sliding window scanning the image, extracting a few statistics, is enough for us to judge "is this surface smooth? Plastic or metal? Will it reflect?"

For the AI community, this is a reminder: pursuing physical realism is not necessarily the right path to human-level perception. Sometimes the deeper you go (deeper networks, more complex physical modeling), the further you get from the target.

For interpretable AI, it is encouraging: when a model is simple enough, explanation is no longer an afterthought. You can directly visualize the weights of a 15×15 kernel and see what kind of local image structure it responds to.

Reference: Morimoto, T., Akbarinia, A., Storrs, K. R., Cheeseman, J. R., Smithson, H. E., Gegenfurtner, K. R., & Fleming, R. W. (2025). Human gloss perception reproduced by tiny neural networks. *bioRxiv*, 2025.05.09.653112. https://doi.org/10.1101/2025.05.09.653112