LRLG： Depth Image Inpainting: Improving Low Rank Completion with Low Gradient Regularization

Table of Contents

• Introduction
• Performance
  • Datasets
  • Compared Algorithms
  • Results
• Building Instruction
  • Prerequisites
  • Compile
• Usage
• Reference
• License

Introduction

LRLG is an algorithm for single depth image inpainting.

Performance

Datasets

• Depth Inpainting Dataset

Compared Algorithms

• [LR] (https://epubs.siam.org/doi/abs/10.1137/080738970)
• [LRTV] (https://ieeexplore.ieee.org/abstract/document/7113897/)
• LRL0: a designed based-line, which employs L0 gradient minimization.

Results

LRLG achieved the best search performance among all the compared algorithms.

Building Instruction

Prerequisites

• GCC 4.5+
• CMake 2.8+
• OpenCV 2.4.1+

Compile

1. Install Dependencies:

$ sudo apt-get install g++ cmake opencv-dev

2. Compile:

$ git clone https://github.com/ZJULearning/depthInpainting.git
$ mkdir build/ && cd build/
$ cmake ..
$ make -j

Usage

TV norm: ./depthInpainting TV depthImage

PSNR calc: ./depthInpainting P depthImage mask inpainted

Inpainting: ./depthInpainting LRTV depthImage mask outputPath"

Generating: ./depthInpainting G depthImage missingRate outputMask outputMissing

LowRank: ./depthInpainting L depthImahe mask outputpath

LRTVPHI: ./depthInpainting LRTVPHI depthImage mask outputPath

TVPHI norm: ./depthInpainting TVPHI depthImage

LRL0: ./depthInpainting LRL0 depthImage mask outputPath initImage K lambda_L0 MaxIterCnt

LRL0PHI: ./depthInpainting LRL0PHI depthImage mask outputPath initImage K lambda_L0 MaxIterCnt

L0: /depthInpainting L0 depthImage

Reference

Reference to cite when you use LRLG in a research paper:

@article{xue2017depth,
  title={Depth image inpainting: Improving low rank matrix completion with low gradient regularization},
  author={Xue, Hongyang and Zhang, Shengming and Cai, Deng},
  journal={IEEE Transactions on Image Processing},
  volume={26},
  number={9},
  pages={4311--4320},
  year={2017},
  publisher={IEEE}
}

License

LRLG is MIT-licensed.