A Single 15×15 Convolution Kernel Reproduces Human Gloss Perception, Challenging the Inverse Physics Hypothesis

Forum topic · 小凯 · 2026-05-25

Summary

A study by researchers from the University of Oxford, Justus Liebig University Giessen, the University of Sydney, and the University of Aberdeen challenges the long-standing 'inverse physics' hypothesis that human gloss perception requires the brain to infer surface reflectance properties from retinal images. Using 3,888 rendered images combining 36 lighting environments, 36 object geometries, and 3 viewpoints, the team collected human gloss judgments via crowd-sourced experiments. Surprisingly, a tiny convolutional neural network containing only a single 15×15-pixel kernel reached 75.3% of the human inter-observer consistency ceiling, while ResNet18 (~11M parameters) performed poorly at predicting human perception. The findings suggest that human vision relies on simple, non-physical image statistics—such as local luminance distributions and contrast patterns—rather than solving physical equations involving BRDFs or lighting estimation. The work, published on bioRxiv (doi:10.1101/2025.05.09.653112), has implications for interpretable AI and low-compute vision applications, including mobile material recognition and edge-device surface inspection, and suggests visual cortex neurons may extract local image statistics rather than perform complex physical inference.

> 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 human perception of object "gloss" requires the brain to perform complex inverse physics computations—inferring 3D surface reflectance properties from 2D retinal images. This story has been told for decades: light hits a surface, the surface reflects it, and the brain "solves" the physical equation.

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

They found: the human visual system judges gloss not through physical inference, but through minimal, non-physical image features. No deep network like ResNet18 pursuing physically faithful representations is needed—a micro-network containing just a single 15×15-pixel convolution kernel achieves 75.3% of the human inter-observer consistency ceiling.

2. Experimental Design: A Systematic Probe with 3,888 Images

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—from simple to complex, familiar to novel.

2.2 Human Blind Test Data

A large-scale online crowd-sourcing experiment collected human judgments to establish a benchmark consistency ceiling for human gloss perception. This is not a "correct answer" but a measure of "how consistent humans are"—the agreement between two people judging the glossiness of the same image.

3. Model Comparison: Smaller Is More Human-Like

3.1 Contending Models

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

3.2 Core Results

ResNet18 failed badly. This deep network, a star on ImageNet, performed poorly at predicting human perception. Its "physically correct" features are seriously misaligned with the judgment strategies humans actually use.

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

Reached 75.3% of the human consistency ceiling
Has a negligible parameter count
Is fully interpretable—the kernel's 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. But this research demonstrates that the brain:

> Does not need to know BRDFs (bidirectional reflectance distribution functions), not need to estimate lighting direction, not need to solve any physical equations.

It only needs to detect simple, generic image statistics—such as skewness of the luminance distribution, contrast patterns, or local structure within a 15×15 neighborhood. These features have no direct physical correspondence, yet they 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 "more compute wins" but evidence that "less is more."

ResNet18's 11 million parameters learn "physically correct features," not "the features humans use"
The 15×15 kernel's 225 parameters 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 for robots

4.3 Implications for Neuroscience

If the brain uses such a simple algorithm to judge gloss, neurons in visual areas V1/V2 may not be performing complex physical inference, but rather extracting local image statistics.

This continues the Fleming lab's earlier line of work: in 2021, their *Journal of Vision* paper "Gloss perception: Searching for a deep neural network that behaves like humans" found that shallow 3–6-layer CNNs better simulated human gloss perception than deeper networks. This new study pushes the conclusion to the extreme—one layer, one kernel is enough.

5. Limitations and Extensions

Limitations: 1. The 3,888 images cover the parameter combination space but are still synthetic renderings; generalization to natural scenes remains to be verified 2. The 15×15 kernel performs well on simple features, but complex materials (e.g., translucency, multi-layer coatings) 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 *Nature Human Behaviour* (2021) paper, "Unsupervised learning predicts human perception and misperception of gloss," showed unsupervised learning can predict human perceptions and illusions
The new study further lowers the floor on the model scale "sufficient to simulate humans"
The direction is clear: minimal, interpretable, biologically inspired models may align better with cognitive reality than blindly stacking depth

6. Verdict: A Conclusion Almost Unsettling in Its Simplicity

The study's conclusion is simple to the point of being provocative:

> 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 sweeping across an image, extracting a few statistics, is enough for us to judge "is this surface glossy? Plastic or metal? Reflective?"

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

And for "interpretable AI," it is encouraging: when a model is simple enough, explanation is no longer an afterthought. With a 15×15 kernel, you can directly visualize the learned weights and see what local image structures 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