Incorporating the loss function into discriminative clustering of structured outputs