Semantic View Synthesis

Ablation study - Background inpainting.

We explore alternative methods for handling the dis-occluded regions when rendering at novel views. We use the standard backward warping to project the synthesized color image using disparity image to render the novel views. We then inpaint the missing pixels using different existing methods.

(a) None: foreground only

(b) Diffusion: simple diffusion (implemented in OpenCV)

(c) GatedConv: a learning-based image inpainting model [Yu et al. ICCV 2019]

(d) Cascade (KB) [Niklaus et al. TOG 2019]

(e) Ours

Note that the depth map used by KB is different from our inferred depth from labels.