Deep-Light-Field-Driven-Saliency-Detection-from-A-Single-View

Introduction

Accepted paper in IJCAI2019, 'Deep Light-Field-Driven Saliency Detection from A Single View', Yongri Piao, Zhengkun Rong, Miao Zhang, Xiao Li and Huchuan Lu.

Dataset: DUTLF

• This dataset consists of DUTLF-MV, DUTLF-FS and DUTLF-Depth.
• The dataset will be expanded to 3000 or so real scenes.
• We are working on it and will make it publicly available soon.

Dataset: DUTLF-MV

• DUTLF-MV is part of DUTLF, which consists of 1580 real scenes.
• Each scene of this dataset consists of an all-focus image, multi-view images and a corresponding ground truth.
• Dataset can be downloaded from here. Code: 9c7k
• Training set: 1100 samples before data augmentation
• Testing set: 480 samples

Usage Instructions

Requirements

• Windows 10
• Tensorflow 1.10.0
• CUDA 9.0
• Cudnn 9.0
• Python 3.6.5
• Numpy 1.14.3

Training

• Download pretrained vgg-19.npy from here. Code: yiov
• Hyperparameter: is_training=1
• Modify your path of training dataset
• Run Main_model
• cd 'your path'/logs, tensorboard --logdir=train

Testing

• Download pretrained model from here. Code: eu72
• Hyperparameter: is_training=0
• Modify your path of testing dataset
• Run Main_model to generate saliency maps, synthesized mutli-view images and depth maps

Saliency map

Saliency maps of this paper can be downloaded BaiduYun. Code: 2jl0

Contact and Questions

Contact: Zhengkun Rong. Email: [email protected] or [email protected]