The authors extend the work of [Schwing et al., Efficient Structured Prediction with Latent Variables for General Graphical Models; ICML 2012] to include active learning protocols. In essence, they use the entropy of the local variable marginals to estimate a value of uncertainty -- so this can be viewed as local variable uncertainty sampling for structured predictions with partially observed variables. To this end, they propose two active learning variants, namely (1) separate active: each active learning round requires a "separate" inference step over the unlabeled and partially labeled variables after learning on the observed variables and (2) joint active: each active learning round follows a joint learning procedure over the labeled, partially labeled, and unlabeled instances. They also explore batch sizes and warm-starting of the learning procedure between rounds in this setting. The empirical evaluation is 3D layout of rooms (a computer vision problem explored in [Scwing et al., ICML 2012], demonstrating that (1) "joint" active learning works the best, achieving an annotation savings of 90%, (2) batch mode works for reasonable small batches, (3) querying "full" vs. "partial" labels [unsurprisingly, partial labels works better] (4) sensitivity to \epsilon and (5) computational reuse saves time [unsuprisingly].