The paper that tries to disentangle appearance from perspective based on a variational autoencoder with a spatial transformer layer. The paper proposes a novel architecture and, as one of the first empirical works, builds on recent work of linking disentangled representations with symmetry transformations. Several reviewers appreciated the author response that also contained new experiments. The original experiments were based on mostly simplistic datasets, but the authors convincingly provided more experimental results during the rebuttal process. The evaluation could further be improved by adopting other metrics as pointed out by R1.