LoLep: Single-View View Synthesis with Locally-Learned Planes and Self-Attention Occlusion Inference.