Table 1 Different MIL pooling methods for FG localization where, \(\bar {y} = [y_{1}, y_{2},..., y_{N}]\)
Pooling method | Pooling operation | Instance label |
|---|---|---|
Max pooling | y=max(yi) | yi |
Average pooling | \(y = 1/N \sum _{i} y_{i}\) | yi |
Attention (softmax) | \(y = \sum _{i} a_{i} y_{i} a_{i} = softmax(w^{T} f(\bar {y})) f(\bar {y}) = tanh(V \bar {y}^{T})\odot sigmoid(U \bar {y}^{T})\) | aiyi |
Attention (sigmoid) | \(y = 1/N \sum _{i} a_{i} y_{i} a_{i} = sigmoid(w^{T} f(\bar {y}))\) | aiyi |
Hybrid (attention + max pooling) | \(y = max (a_{i} y_{i}) a_{i} = sigmoid(w^{T} f(\bar {y}))\) | aiyi |