GilbertRC / LightFieldAngularSR

Fast Light Field Reconstruction With Deep Coarse-To-Fine Modeling of Spatial-Angular Clues

Code for the ECCV 2018 Paper

Fast Light Field Reconstruction With Deep Coarse-To-Fine Modeling of Spatial-Angular Clues

------------------------------------------------------------------------------------------

Please also read our TIP 2018 paper: "Light Field Spatial Super-resolution Using Deep Efficient Spatial-Angular Separable Convolution" with code below

Pytorch - https://github.com/jingjin25/LFSSR-SAS-PyTorch

Matlab - https://github.com/spatialsr/DeepLightFieldSSR

------------------------------------------------------------------------------------------

Description

A learning based model that generate a densely-sampled LF fast and accurately from a sparsely-sampled LF in one forward pass.

Requirements and Dependencies

• MATLAB
• cuda and cudnn (For GPU. Please modify install.m if not using cudnn)
• matconvnet (Please use the matconvnet code given in this repository. It contains the 4D convolution code written by us)

Installation

# Start MATLAB
$ matlab
>> install

Training

Set the training and validation data directory (opts.test_dir) in init_opts.m. Download the training and validation datasets to the specofoc directories. Make sure that there are enough memory for loading the whole training and validatoin datasets.

>> train

Testing Pretrained Models

Set the testing data directory (opts.test_dir) in init_opts.m

>> test

Testing Your Own Models

>> test_model(name, depth, gpu, saveImg, epoch, len)

• model_name : model name
• depth : model depth
• gpu : GPU ID
• saveImg : Save the HR SAIs if true
• epoch : model epoch to test
• len : controls the size of the sub-lightfield, value depends on GPU memory

Authors of the Paper

Henry W. F. Yeung*, Junhui Hou*, Jie Chen , Yuk Ying Chung and Xiaoming Chen

* Equal Contibutions