YanWQ/MANet

Light field depth estimation

MANet

Oral presentation ICASSP 2020 Paper - MANet: Multi-scale Aggregated Network For Light Field Depth Estimation

Update (2021.02.22) Links for downloading our new large-scale wide-baseline light field dataset are moved to the repository LLF-Net

Table of Contents

• Architecture
• Requirement
• Datasets
• Pretrained models
• Inference
• Results
• Citation

Architecture

Requirement

OS system: Windows or Ubuntu
Software: Python == 3.6.7, Tensorflow == 1.12, Pillow == 5.0.0, Opencv == 3.4.5.20
Hardware: CPU or GPU (Geforce 1080ti, 11GB)

Datasets

• Please download the synthetic light field dataset CVIA-HCI dataset from here.
• Please download the EPFL dataset from here, which is captured by the Lytro ILLUM camera.

Pretrained models

Please download our pretrained model from Google drive, which is used in our paper. You can also use it to infer plenoptic (light field) camera datasets.

Inference

python infer.py --dataset YOUR_DATASET --move_path YOUR_IMAGES_MOVE_PATH

Results

Our depth estimation results can be downloaded from Google drive. Here shows an example of the "Cotton" scene from the CVIA_HCI dataset.

Central view	Depth map

Citation

If you use our code in your research, please cite our paper:

@inproceedings{Li2019MANet,
author={Li, Yan and Zhang, Lu and Wang, Qiong and Lafruit, Gauthier},
title={MANet: Multi-scale aggregated network for light field depth estimation},
year={2020},
booktitle={ICASSP 2020}
}