LFSSR-ATO

PyTorch implementation of CVPR 2020 paper: "Light Field Spatial Super-resolution via Deep Combinatorial Geometry Embedding and Structural Consistency Regularization".

[arXiv]

Requirements

• Python 3.6
• PyTorch 1.1
• Matlab (for training/test data generation)

Dataset

We provide MATLAB code for preparing the training and test data. Please first download light field datasets, and put them into corresponding folders in LFData.

Demo

To reproduce final SR reconstruction results in the paper, run:

python demo_LFSSR.py --model_dir pretrained_models --save_dir results --scale 2 --test_dataset Kalantari --angular_num 7 --save_img 1 --crop 1 --feature_num 64 --layer_num 5 2 2 3 --layer_num_refine 3

To reproduce intermediate all-to-one model results in the paper, run:

python demo_ATO.py --model_dir pretrained_models --save_dir results --scale 2 --test_dataset Kalantari --angular_num 7 --save_img 1 --crop 0 --feature_num 64 --layer_num 5 2 2 3

Training

To train the all-to-one model, run:

python train_ATO.py --dataset all --scale 2  --angular_num 7 --feature_num 64 --layer_num 5 2 2 3 --lr 1e-4
python train_ATO.py --dataset all --scale 4  --angular_num 7 --feature_num 64 --layer_num 5 2 2 3 --lr 1e-5

To train the final SR model, run:

python train_LFSSR.py --dataset all --ATO_path pretrained_models/ATONet_2x.pth  --scale 2  --angular_num 7 --feature_num 64 --layer_num 5 2 2 3 --layer_num_refine 3 --weight_epi 0.1  --lr 1e-4