CONTA: Causal Intervention for Weakly-Supervised Semantic Segmentation

As the part of theory is hard to understand, I will start from method part and complete theory part in the future

method

The most difficult part is the meaning of C: confounder set. The $c_i$ is the average of all $X_M$ which is the processed CAMs. I am not confident to this.

The $\alpha$ is a measurement of similarity, whereby get the weight of each class-specific entry c. The $W_1$ and $W_2$ is two learnable matrix.

Another question is how M take part in the calculate of network.

where $s_i=f(X,M_t,{\theta }_t^i)$, f consist of a shared convolutional network and a class-specific fully-connected network. The input of the first part is concat(X, $M_t$) on channel-wise.