This code accompanies the paper
Stereo Magnification: Learning View Synthesis using Multiplane Images
Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, Noah Snavely
SIGGRAPH 2018
Please note that this is not an officially supported Google product.
The entry python script for training is train.py. The input flags are
specified in two places: 1) train.py and 2) stereomag/loader.py.
The input flag which_color_pred specifies how to predict the color image at
each MPI plane:
bg [default] - Our default model. The network predicts: 1) weights for blending
the background and foreground (reference source image) color images at each
plane, 2) the alphas at each plane. 3) a background color image
fgbg - Instead of using the reference source as the foreground image, the
network predicts an extra foreground image for blending with the background
alpha_only - No color image (or blending weights) is predicted by the network.
The reference source image is used as the color image at each MPI plane.
single - The network predicts a single color image shared for all MPI planes.
all - The network directly outputs the color image at each MPI plane.
You can also specify which loss to use for training: pixel or vgg (i.e., the
perceptual loss as measured by differences in VGG features). Note that when
using the VGG loss, you will need to download the pre-trained VGG model
imagenet-vgg-verydeep-19.mat available at
http://www.vlfeat.org/matconvnet/pretrained/#downloading-the-pre-trained-models
The path to this file can be set by the vgg_model_file flag in train.py.
The entry python script for testing the models is test.py.
One could specify what to output to disk by concatenating one or more of the
following (e.g. with '_'): src_images, ref_image, tgt_image, psv, fgbg, poses,
intrinsics, blend_weights, rgba_layers.
psv - the plane sweep volume used as input to the network.
fgbg - foreground and background color images (only valid when
which_color_pred is either fgbg or bg)
blend_weights - weights for blending foreground and backgroud color images (only
valid when which_color_pred is either fgbg or bg)
evaluate.py contains sample code for evaluating the view synthesis performance
based on the SSIM and PSNR metrics. It assumes that each scene result folder
contains a ground-truth target image tgt_image_*.png and the synthesized image
output_image_*.png. The script will output a text file summarizing the metrics
inside the folder FLAGS.result_root.
To run a trained model on a single image pair to generate an MPI, use
mpi_from_images.py. This tool assumes images with the same orientation (as
with a rectified stereo pair), but allows for specifying the (x, y, z) offset
between the images.
You can find an example input stereo pair and command lines for generating results
in the examples directory.
For reference, you can find additional example input stereo pairs, as well as corresponding output multi-plane images and view synthesis results used in the paper in this Google drive link (772 MB).
Our pre-trained model can be downloaded into the models subdirectory by
running the script bash scripts/download_model.sh.
We have released the RealEstate10K dataset suitable for training and testing the MPI model. Note that due to data restrictions, this is not the same version used in our SIGGRAPH'18 paper. However, we are working on updating the results using this public version.