Explicitly Modeled Attention Maps for Image Classification

Self-attention networks have shown remarkable progress in computer vision tasks such as image classification. The main benefit of the self-attention mechanism is ability to capture long-range feature interactions attention-maps. However, computation attention-maps requires a learnable key, query, and positional encoding, whose usage often not intuitive computationally expensive. To mitigate this problem, we propose novel module with explicitly modeled using only single parameter for low computational overhead. design geometric prior based on observation that spatial context given pixel within an mostly dominated by its neighbors, while more distant pixels minor contribution. Concretely, are parametrized via simple functions (e.g., Gaussian kernel) radius, which independently input content. Our evaluation shows our method achieves accuracy improvement up 2.2% over ResNet-baselines ImageNet ILSVRC outperforms other methods AA-ResNet152 0.9% 6.4% fewer parameters 6.7% GFLOPs. This result empirically indicates value incorporating into when applied