Reproduce “Leveraging Geometric Visual Illusions as Perceptual Inductive Biases for Vision Models” on OpenResearch

About this paper

Haobo Yang, Minghao Guo, Dequan Yang, Wenyu Wang

This research integrates geometric visual illusions as a form of perceptual inductive bias into vision models, demonstrating that training with a synthetic dataset of these illusions alongside natural images improves ImageNet classification accuracy, particularly for Vision Transformers, and enhances sensitivity to fine-grained visual structures. The multi-source learning approach achieved gains of up to +0.35 percentage points in Top-1 accuracy for ViT/16 models.

Key insights

• Perception-driven auxiliary tasks, when carefully designed, can systematically improve rather than hinder the primary classification performance of vision models.
• Vision Transformers (ViTs), which inherently lack strong hard-wired locality priors, derived more substantial benefits from geometric illusion supervision compared to ResNet-50, enhancing their structural sensitivity.
• The effectiveness of illusion-induced inductive bias is dependent on factors like image resolution and illusion parameters, suggesting these can be tuned for optimal perceptual transfer.

Summary by alphaXiv.