Baboon MMB

Overview

This repository contains the MATLAB implementation of the Motion Modeling Baseline (MMB) described in the paper "Detecting and Tracking Small and Dense Moving Objects in Satellite Videos: A Benchmark" by Qian Yin et al.

Please note that this is not the original implementation by the authors, as their code was not publicly available.

Getting Started

Prerequisites

• MATLAB
  • Computer Vision Toolbox
  • Global Optimization Toolbox
  • Image Processing Toolbox
  • Optimization Toolbox
  • Parallel Computing Toolbox

Running the Code

Baboon MMB

To run the project with the default parameters, use the following command in the MATLAB console:

baboon_mmb('IMAGE_SEQUENCE', 'path/to/image/sequence/', 'FRAME_RATE', framerate)

This will create an output/ directory which contains:

• amfd/, amfdMasks.mat: Binary masks outputted by the Accumulative Multi-Frame Difference (AMFD) module.
• lrmc/, lrmcMasks.mat: Binary masks outputted by the Low Rank Matrix Completion (LRMC) module.
• combined/, combinedMasks.mat: Binary masks generated by taking the BITWISE_AND (or BITWISE_OR) of the outputs of the AMFD and LRMC modules.
• objects.txt, objects.mat: A struct of all detected objects. Each object contains frameNumber, id, x, y, width, height.
• frames/: The detected bounding boxes drawn over the original image sequence.

Configuration

You can configure the behavior of the algorithms by setting different command-line arguments. The following table lists the parameters you can adjust:

Parameter	Default	Range	Description
K	4	[0, Inf)	Controls the thresholding in the AMFD algorithm. It influences the sensitivity of detection, where a higher value makes the detection more selective.
CONNECTIVITY	8	{4, 8}	Defines the connectivity used in morphological operations for the AMFD algorithm. It can be either 4 or 8, determining how pixel connectivity is considered.
AREA_MIN	5	[1, width * height]	The minimum area threshold for detected regions in the AMFD algorithm. Regions smaller than this value are ignored.
AREA_MAX	80	[1, width * height]	The maximum area threshold for detected regions in the AMFD algorithm. Regions larger than this value are ignored.
ASPECT_RATIO_MIN	1	[1, max(width, height)]	The minimum aspect ratio for detected bounding boxes in the AMFD algorithm. Aspect ratios smaller than this value are ignored.
ASPECT_RATIO_MAX	6	[1, max(width, height)]	The maximum aspect ratio for detected bounding boxes in the AMFD algorithm. Aspect ratios larger than this value are ignored.
L	4	[0, total_seconds]	The length of frames considered for the LRMC algorithm. This parameter controls how many frames are processed together to detect changes.
MAX_NITER_PARAM	10	[1, Inf)	The maximum number of iterations for the LRMC algorithm to converge.
GAMMA1_PARAM	0.3	[0, 1]	A parameter that controls the regularization strength in the LRMC algorithm.
GAMMA2_PARAM	0.8	[0, 1]	Another parameter that controls the regularization strength in the LRMC algorithm.
KERNEL	3	[0, max(width, height)]	The size of the structuring element used in morphological operations in both AMFD and LRMC algorithms.
BITWISE_OR	false	{true, false}	A boolean parameter that, if true, combines the masks from the AMFD and LRMC algorithms using a bitwise OR operation. If false, it combines them using a bitwise AND operation.
H	3	[0, frameCount - 1]	The minimum number of consistent object detections required across the pipeline for the object to be considered valid in the pipeline filter (PF) algorithm.
PIPELINE_LENGTH	5	[0, frameCount - 1]	The number of frames to consider in the pipeline PF algorithm. This parameter controls the length of the buffer used for object tracking.
PIPELINE_SIZE	7	[0, frame_diagonal]	The maximum distance (in pixels) allowed between object detections in consecutive frames for them to be considered the same object in the PF algorithm.
FRAME_RATE	10	[1, Inf)	The frame rate of the input image sequence, used for temporal processing in the LRMC algorithm.
IMAGE_SEQUENCE	''	N/A	The path to the folder containing the input image sequence. The images are loaded and processed in the sequence.
DEBUG	true	{true, false}	A boolean parameter that, if true, enables saving intermediate results and additional debugging outputs.

Optimize

If you would prefer to iteratively run baboon_mmb() to determine the best hyperparameters, optimize.m can be configured and compiled. Note this operation is long and is intended to be run on a cloud computer.

To use the optimize function, you need to configure the config.json file. Here is an example of the config.json file:

{
  "lb": [
    "0",
    "1",
    "1",
    "1",
    "1",
    "1",
    "0",
    "0",
    "1",
    "0",
    "0",
    "0",
    "1",
    "0",
    "0"
  ],
  "ub": [
    "Inf",
    "2",
    "Inf",
    "Inf",
    "Inf",
    "Inf",
    "Inf",
    "Inf",
    "2",
    "Inf",
    "Inf",
    "Inf",
    "Inf",
    "1",
    "1"
  ],
  "mu": [
    "4",
    "2",
    "5",
    "80",
    "1",
    "6",
    "4",
    "3",
    "1",
    "5",
    "7",
    "3",
    "10",
    "0.3",
    "0.8"
  ],
  "std": [
    "1",
    "0.25",
    "1",
    "19.75",
    "0.25",
    "1.25",
    "1",
    "0.75",
    "0.25",
    "1.25",
    "1.75",
    "0.75",
    "2.25",
    "0.075",
    "0.05"
  ],
  "intIndices": [
    2,
    3,
    4,
    8,
    9,
    10,
    11,
    12,
    13
  ],
  "InputPath": "input/viso_video_1",
  "GroundTruthPath": "input/viso_video_1_gt.txt",
  "FrameRate": "10",
  "PopulationSize": "1000",
  "MaxGenerations": "1e9",
  "FunctionTolerance": "1e-10",
  "MaxStallGenerations": "1e6",
  "UseParallel": "false",
  "ParetoFraction": "0.7",
  "Display": "iter",
}

Citations

If you use this code in your research, please cite the following papers:

@article{yin2021detecting,
      title={Detecting and Tracking Small and Dense Moving Objects in Satellite Videos: A Benchmark},
      author={Yin, Qian and Hu, Qingyong and Liu, Hao and Zhang, Feng and Wang, Yingqian and Lin, Zaiping and An, Wei and Guo, Yulan},
      journal={IEEE Transactions on Geoscience and Remote Sensing},
      year={2021},
      publisher={IEEE}
    }

@inproceedings{rezaei2017background,
        title={Background Subtraction via Fast Robust Matrix Completion},
        author={Rezaei, Behnaz and Ostadabbas, Sarah},
        booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
        pages={1871--1879},
        year={2017}
    }