MIMRF:

Multi-Resolution Multi-Modal Sensor Fusion For Remote Sensing Data With Label Uncertainty

Xiaoxiao Du and Alina Zare

If you use this code, cite it: Xiaoxiao Du & Alina Zare. (2019, April 12). GatorSense/MIMRF: Initial Release (Version v1.0). Zenodo. http://doi.org/10.5281/zenodo.2638382

[arXiv] [BibTeX]

In this repository, we provide the papers and code for the Multiple Instance Multi-Resolution Fusion (MIMRF) Algorithm.

Installation Prerequisites

This code uses MATLAB Statistics and Machine Learning Toolbox, MATLAB Optimization Toolbox and MATLAB Parallel Computing Toolbox.

Demo

Run demo_main.m in MATLAB.

Main Functions

The MIMRF Algorithm runs using the following function:

[measure,initialMeasure, Analysis] = learnCIMeasure_minmax_multires(Bags, Labels, Parameters,trueInitMeasure)

Inputs

#The TrainBags input is a 1xNumTrainBags cell. Inside each cell, NumPntsInBag x nSources cell. Inside each cell, the "collection" of all possible combinations generated from the multi-resolution data set. Details please see Section 3 of the MIMRF paper.

#The TrainLabels input is a 1xNumTrainBags double vector that takes values of "1" and "0" for two-class classfication problems -- Training labels for each bag.

Parameters

The parameters can be set in the following function:

[Parameters] = learnCIMeasureParams();

The parameters is a MATLAB structure with the following fields:

1. nPop: size of population
2. sigma: sigma of Gaussians in fitness function
3. maxIterations: maximum number of iterations
4. fitnessThresh: fitness threshold
5. eta: percentage of time to make small-scale mutation
6. sampleVar: variance around sample mean
7. mean: mean of CI in fitness function. This value is always set to 1 (or very close to 1) if the positive label is "1".
8. analysis: if ="1", save all intermediate results

Parameters can be modified by users in [Parameters] = learnCIMeasureParams() function.

Inventory

• Note: some of the util functions were also used in our MICI algorithm. Check out the repo here: [MICI Repository]

• Note: the CI-QP (CI fusion using quadratic programming) approach was also implemented and available in our MICI repository.

https://github.com/GatorSense/MIMRF

└── root dir
    ├── demo_main.m   //Run this. Main demo file.
    ├── demo_MultiRes_data_MU.mat //Demo multi-resolution dataset
    ├── generateSimData_MU.m //Generates synthetic five-source multi-resolution data
    ├── learnCIMeasureParams.m  //parameters function
    ├── MIMRF_Paper.pdf  //related publication
    ├── learnCIMeasure_minmax_multires.m //MIMRF fusion learning function (learns a fuzzy measure from bag-level data and labels)
    ├── evalFitness_minmax_multires.m //MIMRF fusion fitness evaluation
    ├── computeTestMap.m //MIMRF fusion stage (after learning the optimal fuzzy measure g*, compute CI fusion results)
    └── util  //utility functions
        ├── ChoquetIntegral_g_MultiSources.m  //compute CI for multiple sources
        ├── computeci.m    //compute CI fusion output
        ├── evalInterval.m    //evaluate valid intervals of a fuzzy measure
        ├── ismember_findrow_mex.c  //find row index if vector A is part of a row in vector B.   *Need to run "mex ismember_findrow_mex.c"*
        ├── ismember_findrow_mex_my.m  // find row index if vector A is part of a row in vector B (uses above c code).
        ├── share.h  //global variable header to be used in computeci.c
        ├── invcdf_TruncatedGaussian.m //compute inverse cdf for Truncated Gaussian
        ├── rowcol.m //compute row and column index given image index
        ├── sampleMeasure.m //sample new measures
        ├── sampleMeasure_Above.m  //sampling a new measure from top-down.
        ├── sampleMeasure_Bottom.m  //sampling a new measure from bottom-up.
        └── sampleMultinomial_mat.m  //sample from a multinomial distribution.

License

This source code is licensed under the license found in the LICENSE file in the root directory of this source tree.

This product is Copyright (c) 2018 X. Du and A. Zare. All rights reserved.

Citing MIMRF

If you use the MIMRF multi-resolution fusion algorithm, please cite the following reference using the following BibTeX entries.

@article{du2018multi,
  title={Multi-Resolution Multi-Modal Sensor Fusion For Remote Sensing Data With Label Uncertainty},
  author={Du, Xiaoxiao and Zare, Alina},
  journal={arXiv preprint arXiv:1805.00930},
  year={2018}
}

Related Work

Also check out our MICI (Multiple Instance Choquet Integral) algorithm for classifier fusion and regression!

[IEEEXplore (MICI Classifier Fusion and Regression paper)]

[GitHub Code Repository]