Comparison with Soft3D and Zhou et al.

Here we show rendered views from Soft3D, Zhou⁺ (an adaptation of the method described in [1]), and from our method on the Spaces dataset, using the 4-view (large baseline) setup.

As discussed in the paper, Zhou⁺ used 40 planes, increasing the planes further was not possible, The DeepView results below were generated with the 80 plane version of the model.

Hover over the labels at the left to select which output to view. Hold mouse down anywhere to flip to ground truth for comparison.

[1] T. Zhou, R. Tucker, J. Flynn, G. Fyffe, and N. Snavely. Stereo magnification: Learning view synthesis using multiplane images. In Proc. SIGGRAPH, 2018. In SIGGRAPH 2018

Ground truth

Soft3D

DeepView

Zhou⁺

DeepView (depth)

Zhou⁺ (depth)

scene_000

scene_009

scene_010

scene_023

scene_024

scene_052

scene_056

scene_062

scene_063

scene_073