Seonglae ChoFoveal vision, Foveation , Foveated sampling
Resolution difference between center and periphery due to cell density
Seeing more with less: human-like representations in vision models
Human vision's foveation approach (high-resolution center, low-resolution periphery) is applied directly to evaluate existing vision and multimodal models.
With the same pixel budget, foveated sampling outperforms uniform sampling in accuracy: up to ~2–3%p improvement in VQA and object detection. Only 3% of pixels achieve ~80% of full performance → demonstrating Diminishing returns with increasing resolution. Internal analysis shows that foveation induces human-like representations including more global self-attention and resolution-specific neuron selectivity.