Exploit and Replace: An Asymmetrical Two-Stream Architecture for Versatile Light Field Saliency Detection

Introduction

Accepted paper in AAAI2020, 'Exploit and Replace: An Asymmetrical Two-Stream Architecture for Versatile Light Field Saliency Detection', Yongri Piao, Zhengkun Rong, Miao Zhang and Huchuan Lu.

MATLAB Code for Stacking Focal

We update MATLAB script for stacking focal slices on September 2, 2021. The generation of .mat files (stacked focal slices) was implemented by a MATLAB script. To be more specific, we first read the focal slices (12 focal slices in a scene ordered by the shallowest depth first) and stacked them according to their depth order, then generated a .mat file for each scene.

Usage Instructions

Requirements

• Windows 10
• PyTorch 0.4.1
• CUDA 9.0
• Cudnn 7.6.0
• Python 3.6.5
• Numpy 1.16.4

Exploit -- Focal Stream (Teacher)

Training

• Modify your path of training dataset in Demo_Teacher
• Set args.phase = train
• Set args.param = False
• Run Demo_Teacher

Testing

• Download pretrained focal model from here. Code: vee3
• Modify your path of testing dataset in Demo_Teacher
• Set args.phase = test
• Set args.param = True
• Run Demo_Teacher to inference saliency maps

Replace -- RGB Stream (Student)

Training

• Modify your path of training dataset in Demo_Student
• Set args.phase = train
• Set args.param = False
• Run Demo_Student

Testing

• Download pretrained RGB model (comming soon)
• Modify your path of testing dataset in Demo_Student
• Set args.phase = test
• Set args.param = True
• Run Demo_Student to inference saliency maps

Saliency Maps

Focal Stream (Teacher)

• Download Link. Code: 58to

RGB Stream (Student)

• Download Link. Code: nfqs

Contact and Questions

Contact: Zhengkun Rong. Email: 18642840242@163.com or rzk911113@mail.dlut.edu.cn