We experimentally analyze learning structured output in a discriminative framework where values of the output variables are estimated by local classifiers. In this framework, complex dependencies among the output variables are captured by constraints that dictate how global labels can be inferred. We compare two strategies, learning plus inference and inference based training, by observing their behaviors in different conditions. We conclude that using inference during learning helps when the local classifiers are difficult to learn but requires more examples.