Learning to complete partial observations from unpaired prior knowledge

Highlights：

• A novel single-step training strategy for the hallucinating task that simultaneously utilizes knowledge from a partially observed domain without ground truth and an unpaired prior knowledge domain.

• In this setting, we are the first to benchmark using two independent hallucinating tasks, i.e., 2-D road layout hallucinating and 3-D vehicle shape completion.

• Using these benchmarks, we demonstrate that our proposed training strategy outperforms current state-of-the-arts, i.e., a variational auto-encoder based approach and a adversarial training based approach, in terms of level-of-detail and generalization for unseen data.