DeepView
View Synthesis with Learned Gradient Descent

Technical Video

Video with synthesized fly-throughs and depth visualizations of the scenes shown in the paper.

Example MPIs in our interactive viewer

We present several scenes in an interactive viewer.
Note that these were made with a 16-view version of the model in the paper with a sparsity penalty to reduce unneeded content on occluded layers.
The Chrome browser is recommended.

Scene 9
low | high

Scene 10
low | high

Scene 23
low | high

Scene 52
low | high

Scene 56
low | high

Scene 63
low | high

Here are brief instructions for using the viewer.

Comparison with Soft3D and Zhou et al. on Spaces dataset

View a comparison on test scenes for 4-view (large baseline) case.

Comparison with Soft3D on Kalantari et al. dataset

View a comparison on test scenes from Kalantari et al.

Other Studies

Click to see results from ablation of gradient components, or from varying the number of LGD iterations.

Extended Training Details

Click to see details of the methods used to reduce RAM during training and inference, as well as the training and loss hyperparameters.

Spaces training data

Click here to access the Spaces dataset used to train DeepView and a script to compute the evaluation in the paper.

VR@50 2018 Light Fields

At SIGGRAPH 2018 in Vancouver, Prof. Henry Fuchs invited us to record panoramic light field stills of the "VR@50" panels featuring Virtual Reality pioneer Ivan Sutherland and his colleagues. Here you can find the light field data files which can be viewed in 6-degrees-of-freedom VR using the free Welcome to Light Fields app available on Steam VR. First, here are 4K panoramas rendered from the center position of the light fields of both the afternoon and the evening events:

Installing the light fields will allow you to step inside these scenes in Virtual Reality and move your viewpoint in any direction in about a 60 centimeter diameter volume. Download the additional VR@50 light fields and installation instructions here. Enjoy!