Train Cascade Object Detector bounding box out of bound?

Question

I am trying to train a cascade object detector using the built-in function in Matlab (vision toolbox). However, the following message came up after running the command.

*

Error using trainCascadeObjectDetector (line 245)
Error reading instance 1 from image 2, bounding box possibly out of image bounds.

*

I don't understand why the bounding box can be out of bounds. All the parameters for my positive images are set up correctly (starting point x,y, width, and height. I used createMask(h) to create a mask and find the minimum coordinates for x and y to be starting point and max-min for each dimension to be the width and height), and the negative images (as far as I know) are just images without any setup needed.

Anyone ever ran into the same problem? How did you solve it?

EDIT: Here's the code. I don't have the toolbox for training the "data" struct, so I wrote one myself

positive_samples=struct;
list=dir('my_folder_name_which_I_took_out');

L=length(list)-3; %Set L to be the length of the image list.
for i=1:length(list)
    positive_samples(i).imageFilename=list(i).name;
end
positive_samples(:,1)=[]; %first 3 lines do not contain file names
positive_samples(:,1)=[];
positive_samples(:,1)=[];

for j=1:1
    imshow(positive_samples(j).imageFilename);
    title(positive_samples(j).imageFilename);
    h=imrect;
    h1=createMask(h);
    I=imread(positive_samples(j).imageFilename);
    [le, wi, hi]=size(I);
    tempmat=[];
    count=1;
    for l=1:le
        for m=1:wi
            if h1(l,m)==1
                tempmat(count,1)=l;
                tempmat(count,2)=m;
                count=count+1;
            end
        end
    end
    positive_samples(j).objectBoundingBoxes(1,1)=min(tempmat(:,1));
    positive_samples(j).objectBoundingBoxes(1,2)=min(tempmat(:,2));
    positive_samples(j).objectBoundingBoxes(1,3)=max(tempmat(:,2))-min(tempmat(:,2));
    positive_samples(j).objectBoundingBoxes(1,4)=max(tempmat(:,1))-min(tempmat(:,1));
    imtool close all
end
trainCascadeObjectDetector('animalfinder.xml', positive_samples, 'my_neative_folder_name', 'FalseAlarmRate', 0.2, 'NumCascadeStages', 3);

sorry if it's messy......

Answer 1

I did not run the code, because I don't own the toolbox, but the following lines are very "suspicious":

positive_samples(j).objectBoundingBoxes(1,1)=min(tempmat(:,1));
positive_samples(j).objectBoundingBoxes(1,2)=min(tempmat(:,2));
positive_samples(j).objectBoundingBoxes(1,3)=max(tempmat(:,2))-min(tempmat(:,2));
positive_samples(j).objectBoundingBoxes(1,4)=max(tempmat(:,1))-min(tempmat(:,1));

I would expect:

positive_samples(j).objectBoundingBoxes(1,1)=min(tempmat(:,2));
positive_samples(j).objectBoundingBoxes(1,2)=min(tempmat(:,1));
positive_samples(j).objectBoundingBoxes(1,3)=max(tempmat(:,2))-min(tempmat(:,2));
positive_samples(j).objectBoundingBoxes(1,4)=max(tempmat(:,1))-min(tempmat(:,1));

Some suggestions to shorten your code, they are not related to the problem:

You can shorten line 4 to 9 to a single line, avoiding the loop: [positive_samples(1:L).im]=list(4:end).name

And this loop can be replaced as well:

tempmat=[];
count=1;
for l=1:le
    for m=1:wi
        if h1(l,m)==1
            tempmat(count,1)=l;
            tempmat(count,2)=m;
            count=count+1;
        end
    end
end

shorter and faster code:

[y,x]=find(h1);
tempmat=[y x];

Answer 2

There is a better way to label your positive samples. The Computer Vision System Toolbox now includes the Training Image Labeler app (as of release 2014a). If you do not have R2014a you should try the Cascade Training GUI app.