This project provides the code for 'Enhanced Light Field Reconstruction by Combining Disparity and Texture Information in PSVs via Disparity-Guided Fusion', IEEE TCI, 2023. paper link
Note: The explicit-depth-based and implicit-depth-based pipelines adopt the basic structure of GA-Net and MALFRNet (w/o their refinement), respectively.
- Python=3.7
- PyTorch=1.8.0
- scikit-image=0.14.2
- Matlab (for .h5 file generation)
- Download the TrainingSet (code: 3f2x) and TestSet (code: 6c31) and put them under './LFData/' folder.
- Run
PrepareData_xxx.m
to generate .h5 file for training and test. - Or directly download our generated .h5 file (code: sgca).
model_HCI for synthetic datasets, 2x2→7x7 interpolation
python train_HCI.py --train_dataset HCI --disp_range 4 --num_planes 50 --angular_in 2 --angular_out 7 --epoch 50000 --learning_rate 1e-4 --decay_rate 0.5 --decay_epoch 5000 --batch_size 1 --patch_size 64
model_SIG for real-world datasets, 2x2→7x7 interpolation
python train.py --train_dataset SIG --disp_range 1.5 --num_planes 32 --angular_in 2 --angular_out 7 --epoch 10000 --learning_rate 1e-4 --decay_rate 0.5 --decay_epoch 1000 --batch_size 1 --patch_size 64
The training curve for disp_thres appears like:
Synthetic datasets (HCI, HCI old and Inria DLFD), 2x2→7x7 interpolation
python test_HCI.py --model_dir pretrained_model --train_dataset HCI --disp_range 4 --num_planes 50 --angular_in 2 --angular_out 7 --input_ind 0 6 42 48 --crop 1
Real-world datasets (30scenes, occlusions and reflective), 2x2→7x7 interpolation
python test.py --model_dir pretrained_model --train_dataset SIG --disp_range 1.5 --num_planes 32 --angular_in 2 --angular_out 7 --input_ind 0 6 42 48 --crop 0
Our performance under the 2x2→7x7 interpolation task:
@ARTICLE{LFASR-ELFR,
title={Enhanced Light Field Reconstruction by Combining Disparity and Texture Information in PSVs via Disparity-Guided Fusion},
author={Yilei Chen and Xinpeng Huang and Ping An and Qiang Wu},
journal={IEEE Transactions on Computational Imaging},
year={2023},
volume={9},
pages={665-677},
month={Jul.}}
Any questions regarding this work can contact yileichen@shu.edu.cn.