Computer Vision News - June 2018

= ( 1 , . . . , ) and is the vertex sampled from the model at time t, and = ( ( ), 4. Visual Attention Mechanism Previous RNN hidden state image feature map is used by the RNN to focus only on the relevant information in the next time step. For this, at time step t, the following weighted feature map is computed: = ( ( , 1 (ℎ 1, −1 ), 2 (ℎ 2, −1 ))) = ∘ where ∘ is the Hadamard product, is the skip feature tensor, ℎ 1, and ℎ 2, are the hidden state tensors from the two-layers ConvLSTM. 5. Evaluator Network An evaluator network chooses among a list of K=5 candidate polygons. This network takes as input the skip features (112x112 blue tensor in CNN Encoder figure), the last state tensor of the ConvLSTM, and the predicted polygon, and tries to estimate its quality. In training, we minimize the mean squared error ( ) = [ ( , ) − ( , )] 2 , is the network predicted , is the mask for the sampled vertices and is the ground-truth mask. 6. Gated Graph Neural Network -- GGNN Research 7 Research Computer Vision News