ELSDc: Ellipse and Line Segment Detector, with Continuous validation

V. Patraucean, P. Gurdjos, R. Grompone von Gioi
(Corresponding author: Viorica Patraucean vpatrauc@gmail.com)
Version 0.9, March 2015

ELSDc is an algorithm for joint ellipse and line segment detection in digital images. It is based on the a-contrario framework and should work without the need for parameter tuning. The algorithm is described in the following manuscript:

"Joint A Contrario Ellipse and Line Detection", V. Patraucean, P. Gurdjos, and R. Grompone von Gioi.

Demo

An online demo of the algorithm is available, where you can upload your own images and run ELSDc:

Demo: http://dev.ipol.im/~jirafa/ipol_demo/elsdc/
(user: demo, pass: demo)

Python interface

Available here. Thank you, Patrick!

Requirements

ELSDc requires CLAPACK/CBLAS library for some linear algebra computations. Version 3.2.1 was used.

Compilation

You may need to modify the file Makefile in the directory src to include the path to liblapack. Then, you can compile by typing make in the command line. This produces the executable called elsdc.

Execution

./elsdc image.pgm runs ELSDc on the image specified by image.pgm. This version works only with PGM images. This folder contains the image 'shapes.pgm' for testing purposes.

Output

output.svg contains the execution result in SVG format. shapes_output.svg contains the expected result for the sample image shapes.pgm.
labels.pgm after execution, each pixel in this image is labelled with the label of the primitive to which it belongs. shapes_labels.pgm contains the expected result for the sample image shapes.pgm.
out_ellipse.txt contains the parameters of the detected circular/elliptical arcs in the form label x_c y_c a b theta ang_start ang_end. shapes_out_ellipse.txt contains the expected result for the sample image shapes.pgm.
out_polygon.txt contains the parameters of the detected line segments, grouped into polygons, defined through contact points, in the form label number_of_points x1 y1 x2 y2 x3 y3 .... shapes_out_polygon.txt contains the expected result for the sample image shapes.pgm.

The command will print at the end of the execution the numbers of features of each type found. To check the installation, run ./elsdc shapes.pgm. The output should be similar to the one in shapes_output.svg, shapes_labels.pgm, shapes_out_ellipse.txt and shapes_out_polygon.txt. It should find 66 ellipses and 142 polygons. The exact coordinate values may differ slightly due to rounding error differences in different systems. The execution time for shapes.pgm is about 1.4s on a Dell notebook.

Source Code Files

main.c contains the main() function; entry point into the application, calls IO functions and the detection function.
curve_grow.c contains functions for region grow (gather neighbour pixels that share the same gradient orientation) and curve grow (gather regions that describe a convex and smooth contour).
rectangle.c defines a rectangle structure (i.e. segment with width) to approximate the result of region grow.
polygon.c defines a polygon structure (i.e. collection of rectangles).
ring.c defines a (circular or elliptical) ring structure.
elsdc.c contains refinement and validation functions for different types of primitives (ellipse, circle, line segment).
ellipse_fit.c contains functions to estimate a circle or an ellipse using pixels positions and their gradient orientations.
iterator.c functions to count the number of aligned pixels inside a rectangle.
lapack_wrapper.c contains wrappers for lapack functions for linear system solve.
pgm.c IO functions for pgm image format (the only format supported currently).
image.c defines structures for image representation and functions for gradient computation.
gauss.c defines a Gaussian kernel and contains functions to performs Gaussian filtering of an image.
misc.c contains general-purpose functions and constants definitions.
svg.c functions to write the result in svg format.

Datasets

We make available two datasets together with their ground truth for quantitative evaluation of ellipse detectors:

Dataset1_SyntheticCircles consists of 20 images, 500x500 pixels, containing non-overlapping and overlapping circles. The images are corrupted with five different levels of Gaussian noise, with five different noise realisations for each noise level and for each image, resulting in 500 image instances in total. The file coord.txt contains the ground truth circle parameters in the form: pathToimageName.pgm xcenter1 ycenter1 radius1 xcenter2 ycenter2 radius2 xcenter3 ycenter3 radius3. Ten images of pure Gaussian noise are also added (folder pureNoise), where all detections are false positives.

Dataset2_CalibrationPatterns contains (in folder images) 40 natural images of calibration patterns that were included in Higuchi et al.'s package for camera calibration link. The patterns contain coplanar disjoint and concentric circles. The ground truth was obtained by manually labelling the contours belonging to circles and rings projections. For each image in folder images, there is a corresponding .txt file in folder ground_truth with the same name, containing the primitive parameters in the form: xcenter ycenter major_axis minor_axis theta.

Dataset3_RealImages contains real images that were used to produce the results from the paper "Joint A Contrario Ellipse and Line Detection".

License

This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

viorik/ELSDc