Monocular Neural Image Based Rendering with Continuous View Control

task: real-time monocular novel view synthesis

• w/ continuous viewpoint control
  • previous work overfit to training settings
• input image -> latent code -> transformed latent code -> depth map -> bilinear sampling from input image -> output
• unable to gen. image w/ large viewpoint discrepancy?
• L_1 distance… not useful in my setting at all