Awesome

DI4SLF

PyTorch implementation of ICCV 2021 paper: "Learning Dynamic Interpolation for Extremely Sparse Light Fields with Wide Baselines"

Requrements

• Python 3.7.4
• Pytorch 1.6.0

Train and Test

Data

We provide MATLAB code for preparing the training and test data in the folder ./LFData. We estimate the optical flow of source views using a pre-trained optical-flow model provided by RAFT. Before generating the training and testing datasets, please first put the estimated optical flow of source views in the folder ./LFData/flow_source in the .mat format with the shape [num_lf, an_h, an_w, h, w], where num_lf, [an_h, an_w], [h, w] are the number, angular resolutions, and spatial resolutions of sparse lfs, respectively. Our training dataset can be downloaded from here, and the testing dataset can be downloaded from here.

Test

The testing codes are in the folder ./test

• model: the trained model which is in this subfolder;
• Inference:

python lfr_test.py

Train

The testing codes are in the folder ./train

• Training:

python lfr_train.py