Paper · On the benchmark

A Training-Free Saliency Model

Physics-inspired attention, measured against real gaze

← All research

On MIT/Tübingen Benchmark

Attention as a Magnetic Field: A Biologically-Inspired Saliency Model

Cora Zeng (2026) · Zenodo · DOI: 10.5281/zenodo.18979607

MFA models visual attention as a magnetic field where image features act as sources generating attraction forces. Unlike CNN-based saliency models, MFA uses inverse-square field dynamics inspired by physics, requiring no training data or neural network weights. The model is evaluated on the MIT/Tübingen Saliency Benchmark (MIT300 leaderboard) with AUC 0.7926.

The result that matters: a model with zero learned parameters lands within reach of trained networks. That is evidence the magnetic-field idea is capturing something real about how gaze is drawn, not just curve-fitting — the practical side of the theory described in the companion paper.

visual attention saliency prediction magnetic field biologically inspired eye tracking