# Feret Diameter: Introduction8

Posted by Steve Eddins,

This is the first of a few blog posts about object measurements based on a concept called the Feret diameter, sometimes called the caliper diameter. The diagram below illustrates the concept. Place the object to be measured inside the jaws of a caliper, with the caliper oriented at a specified angle. Close the jaws tightly on the object while maintaining that angle. The distance between the jaws is the Feret diameter at that angle.

Eventually, we'll examine several related measurements (and algorithms for computing them), like these.

Today, though, I'm just going to play around with a very small binary image. (Note: the pixelgrid function is a utility that I wrote. It's at the bottom of this post.)

bw = [ ...
0   0   0   0   0   0   0   0   0   0   0
0   0   0   0   0   0   0   0   0   1   0
0   0   0   0   0   0   1   1   1   1   0
0   0   0   0   0   1   1   1   1   0   0
0   0   0   1   1   1   1   1   1   0   0
0   0   1   1   1   0   0   0   0   0   0
0   0   1   1   1   0   0   0   0   0   0
0   0   0   0   0   0   0   0   0   0   0
0   0   0   0   0   0   0   0   0   0   0 ];

imshow(bw)
pixelgrid


I want to brute-force my to finding the two corners of the white blob that are farthest apart. Let's start by finding all the pixel corners. (I'll be using implicit expansion here, which works in R2016b or later.)

[y,x] = find(bw);
hold on
plot(x,y,'xb','MarkerSize',5)
corners = [x y] + cat(3,[.5 .5],[.5 -.5],[-.5 .5],[-.5 -.5]);
corners = permute(corners,[1 3 2]);
corners = reshape(corners,[],2);
corners = unique(corners,'rows');
plot(corners(:,1),corners(:,2),'sr','MarkerSize',5)
hold off


Well, let's get away from pure brute force just a little. We really only need to look at the points that are part of the convex hull of all the corner points.

hold on
k = convhull(corners);
hull_corners = corners(k,:);
plot(hull_corners(:,1),hull_corners(:,2),'r','LineWidth',3)
plot(hull_corners(:,1),hull_corners(:,2),'ro','MarkerSize',10,'MarkerFaceColor','r')
hold off


Now let's find which two points on the convex hull are farthest apart.

dx = hull_corners(:,1) - hull_corners(:,1)';
dy = hull_corners(:,2) - hull_corners(:,2)';
pairwise_dist = hypot(dx,dy);
[max_dist,j] = max(pairwise_dist(:));
[k1,k2] = ind2sub(size(pairwise_dist),j);

point1 = hull_corners(k1,:);
point2 = hull_corners(k2,:);
hold on
plot([point1(1) point2(1)],[point1(2) point2(2)],'-db','LineWidth',3,'MarkerFaceColor','b')
hold off


And the maximum distance is ...

max_dist

max_dist =

10



(I thought I had made a mistake when I first saw this number come out to be an exact integer. But no, the answer is really 10.)

That's a pretty good start. I don't know exactly where I'll go with this topic next time. I'm making it up as I go.

function hout = pixelgrid(h)
%pixelgrid Superimpose a grid of pixel edges on an image
%   pixelgrid superimposes a grid of pixel edges on the image in the
%   current axes.
%
%   pixelgrid(h_ax) superimposes the grid on the image in the specified
%   axes.
%
%   pixelgrid(h_im) superimposes the grid on the specified image.
%
%   group = pixelgrid(___) returns an hggroup object that contains the two
%   lines that are used to draw the grid. One line is thick and darker, and
%   the other is thin and lighter. Using two contrasting lines in this way
%   guarantees that the grid will be visible regardless of pixel colors.
%
%   EXAMPLES
%
%   Superimpose pixel grid on color image. Zoom in.
%
%       imshow(rgb)
%       pixelgrid
%       axis([440 455 240 250])
%
%   Change the colors and line widths of the pixel grid.
%
%       imshow(rgb)
%       h = pixelgrid;
%       axis([440 455 240 250])
%       h.Children(1).Color = [178 223 138]/255;
%       h.Children(1).LineWidth = 2;
%       h.Children(2).Color = [31 120 180]/255;
%       h.Children(2).LineWidth = 4;
%
%   LIMITATIONS
%
%   This function is intended for use when looking at a zoomed-in image
%   region with a relatively small number of rows and columns. If you use
%   this function on a typical, full-size image without zooming in, the
%   image will not be visible under the grid lines.
%
%   This function attempts to determine if MATLAB is running with a high
%   DPI display. If so, then it displays the pixel grid using thinner
%   lines. However, the method for detecting a high DPI display on works on
%   a Mac when Java AWT is available.

%   Steve Eddins
%   Copyright The MathWorks, Inc. 2017

if nargin < 1
him = findobj(gca,'type','image');
elseif strcmp(h.Type,'axes')
him = findobj(h,'type','image');
elseif strcmp(h.Type,'image')
him = h;
else
error('Invalid graphics object.')
end

if isempty(him)
end

hax = ancestor(him,'axes');

xdata = him.XData;
ydata = him.YData;
[M,N,~] = size(him.CData);

if M > 1
pixel_height = diff(ydata) / (M-1);
else
% Special case. Assume unit height.
pixel_height = 1;
end

if N > 1
pixel_width = diff(xdata) / (N-1);
else
% Special case. Assume unit width.
pixel_width = 1;
end

y_top = ydata(1) - (pixel_height/2);
y_bottom = ydata(2) + (pixel_height/2);
y = linspace(y_top, y_bottom, M+1);

x_left = xdata(1) - (pixel_width/2);
x_right = xdata(2) + (pixel_width/2);
x = linspace(x_left, x_right, N+1);

% Construct xv1 and yv1 to draw all the vertical line segments. Separate
% the line segments by NaN to avoid drawing diagonal line segments from the
% bottom of one line to the top of the next line over.
xv1 = NaN(1,3*numel(x));
xv1(1:3:end) = x;
xv1(2:3:end) = x;
yv1 = repmat([y(1) y(end) NaN], 1, numel(x));

% Construct xv2 and yv2 to draw all the horizontal line segments.
yv2 = NaN(1,3*numel(y));
yv2(1:3:end) = y;
yv2(2:3:end) = y;
xv2 = repmat([x(1) x(end) NaN], 1, numel(y));

% Put all the vertices together so that they can be drawn with a single
% call to line().
xv = [xv1(:) ; xv2(:)];
yv = [yv1(:) ; yv2(:)];

hh = hggroup(hax);
dark_gray = [0.3 0.3 0.3];
light_gray = [0.8 0.8 0.8];

if isHighDPI
bottom_line_width = 1.5;
top_line_width = 0.5;
else
bottom_line_width = 3;
top_line_width = 1;
end

% When creating the lines, use AlignVertexCenters to avoid antialias
% effects that would cause some lines in the grid to appear brighter than
% others.
line(...
'Parent',hh,...
'XData',xv,...
'YData',yv,...
'LineWidth',bottom_line_width,...
'Color',dark_gray,...
'LineStyle','-',...
'AlignVertexCenters','on');
line(...
'Parent',hh,...
'XData',xv,...
'YData',yv,...
'LineWidth',top_line_width,...
'Color',light_gray,...
'LineStyle','-',...
'AlignVertexCenters','on');

% Only return an output if requested.
if nargout > 0
hout = hh;
end

end

function tf = isHighDPI
% Returns true if running with a high DPI default display.
%
% LIMITATION: Only works on a Mac with Java AWT available. In other
% situations, this function always returns false.
tf = false;
if ismac
if usejava('awt')
sf = java.awt.GraphicsEnvironment.getLocalGraphicsEnvironment().getDefaultScreenDevice.getScaleFactor();
tf = sf > 1;
end
end
end


Get the MATLAB code

Published with MATLAB® R2017b

### Note

Cris replied on : 1 of 8
Steve, The Feret diameter is a great topic! I wanted to leave a comment to a post from 2008, but comments are closed there, I hope you don't mind me leaving it here instead. It's about this post: https://blogs.mathworks.com/steve/2008/09/03/dilation-algorithms-introduction/ . If you run that code on R2017a (it's the version I've got installed right now), you'll see a very different plot: the right side is 5 times as high as the left side. You used to have an O(1) algorithm for the dilation, now it is O(n). What happened?
Steve Eddins replied on : 2 of 8
Cris—I'll look into it.
Steve Eddins replied on : 3 of 8
Cris—I took a closer look at the performance curves from this example in both R2008a and R2017b. My first observation is that R2017b is running these imdilate cases a LOT faster - 10x to 20x (on the same computer). My second observation is that the performance curve was never really O(1) (meaning independent of n). In both releases, the curve looks something like K \sqrt{n} + b. So the performance varies with the window width. I'm guessing that's because of a memory allocation associated with an intermediate result. But the overhead term, b, is greatly reduced in R2017b. Here are the performance curves for the two releases on the same plot:
Cris replied on : 4 of 8
Wow, that is a really big speed improvement! Impressive!
Baris Demiroz replied on : 5 of 8
Very nice post, thank you. The use of subpixel coordinates for corners made me wonder if there is an easy way to convert bwboundaries() result to subpixel contour (the boundary between the white pixels and black pixels) instead of pixels. I tried to tackle that problem but I didn't like the solution I came up with, it is too messy. I wonder if there is more elegant way of doing that.
Mark Hayworth replied on : 7 of 8
When is Feret diameter, and largest perpendicular Feret diameter, going to be available via regionprops()? It would be so much more useful that MajorAxisLength and MinorAxisLength.
Steve Eddins replied on : 8 of 8
Mark—I will remind the development team of your request.