Arc Adjacency Matrix based Fast Ellipse Detection

We proposed a fast ellipse detection method based on arc adjacency matrix. We have successfully used this method in some applications, such as satellite tracking, UGV guidance and pose estimation.

:blush:The binaries for Matlab and Python can be downloaded from the latest release.

1 Compile our Codes

We have successfully applied AAMED to various platforms (Windows, Ubuntu, ARM). The codes used for different platforms may require some minor changes.

1.1 Windows

• OpenCV > 3.1.0
• VS 2015

You can add all .h and .cpp files into your project. Don't forget to config your project about OpenCV :).

main.cpp has given an example to detect ellipses from an image.

AAMED aamed(drows, dcols). drows (dcols) must be larger than the rows (cols) of all used images. Then, we can use aamed.run_FLED(imgG); to detect ellipses from multiple images.

Very important: Please check rows and cols of your input images are smaller than drows and dcols separately, otherwise, there will be some errors at class NODE_FC

1.2 Ubuntu

We use CMake to generate Makefile, then use make to compile our method. This way is only used for Ubuntu, not suitable for Windows.

cd AAMED/cmake-build
cmake ..
make 
./AAMED

1.3 Python

For Python, the OpenCV and NumPy packages are required.

Building

With the Anaconda distribution of Python (Windows and Linux) and the standard Python in Linux, building the library can be done in the following way:

cd python
python setup.py build_ext --inplace

Once built, the created library (Windows: .pyd, Linux: .so) can be placed anywhere.

If you are building on Windows without Anaconda, you must install OpenCV manually alongside the OpenCV Python package. (Make sure the versions are the same!) Then, in the setup.py file, the opencv_root variable should be set to the specified OpenCV installation location. Once this is done, you can continue to use the same commands above to build.

Note: for Windows without Anaconda, the opencv_world DLL should be together with the .pyd as well. Alternatively, if you do not want to copy the opencv_world DLL around, you can add the OpenCV bin location as a DLL directory at the beginning of your script. For example:

import os
os.add_dll_directory("D:/opencv/build/x64/vc14/bin")

Testing

To quickly test, test_aamed.py is provided.

python test_aamed.py

1.4 MATLAB

We have packaged AAMED, it can be used in MATLAB. AAMED needs OpenCV support. Note that if mexdestoryaamed(obj) is not called, the memory used in AAMED will remain in MATLAB all the time. Only restart matlab can clear the memory.

Install

You need to config OpenCV include path and library path in setup.m firstly. Then, you can run setup.m to compile mexAAMED, mexdestoryAAMED, mexdetectImagebyAAMED, mexSetAAMEDParameters.

Test

test_aamed.m provides an example to detect ellipses from an image.

obj = mexAAMED(540, 960); % AAMED only needs to be defined once
mexSetAAMEDParameters(obj, pi/3, 3.4, 0.77); % Set the parameters.

% This function can be used multiple times to detect ellipses from images
detElps = mexdetectImagebyAAMED(obj, img); 

mexdestoryAAMED(obj); %Free memory (very important).

2 Label Tool

we provide a tool to label ellipses (circles) from an image. This tool is based on MATLAB R2016. First, you need to run setup.m to compile mexElliFit. Then, you can run main.m to use this label tool.

3 AAMED Viewer

we proivide a tool to show critical data (edge contours, DP contours, arc contours, AAM and detected ellipses) in MATLAB. We use this tool to find bugs of AAMED and test functions.

You need to run setup.m to compile mexcvtBasicData, mexcvtRRect, mexcvtVVP, mexcvtAAM. Then you can use main.m to read DetailAAMED.aamed.

4 Nine Datasets

We have uploaded 9 datasets used in our paper to Baidu Cloud (Code: 7br2) and Google Drive.

5 Measure Tool

We have provided a tool that can be used to measure our method. Firstly, you need to run setup.m in measuretool/MeasureTools to build the mex files. Then, MeasureAllDatasets.m needs to be configured as described below. Finally, you can run MeasureAllDatasets.m to measure the used method.

• data_root_path: The root path of all datasets.
• dataset_name: The file names of used datasets. dataset_name is an array of cell, each element in it is a string that represents the dataset file name.
• gt_label contains the labels of corresponding datasets. Read_Ellipse_GT.m uses these labels to load ground-truth.
• methods_name is the file name that stores the results of used method.
• method_label is the label of corresponding method. Read_Ellipse_Results.m uses this label to load detection results.

The sample output of MeasureAllDatasets.m is as following.

Evaluating dataset: Satellite Images - Dataset Meng #2
Precision: 80.9524%,  Recall: 85%,  F-measure: 82.9268%. 
Average detected time: 2.6065 ms.

6 How to Make your Datasets

The format of ellipse dataset is as follows.

If you want to make a new dataset, you can put the collected images into the file images. Then, the ground-truth files that are generated by labeltool can be put into the file gt. Finally, you need to create the file imagenames.txt that contains all image names.

If you have any questions, please contact me ([email protected]) or create an issue.