Error generating samples for cascade classifier training

While trying to run trainCascadeObjectDetector with MATLAB R2013a on Windows, I keep getting an error that won't allow me to get past the very first stage.

I have been following MATLAB's tutorial on how to "Train a Cascade Object Detector", as well as this blog post .

I am using 4,372 .pgm positive images and 14,440 .pgm negative images.

The positive images are given to the function as a structure variable which contains path+filename and object bounding box of the images. Both positive and negative images are located in their respective subfolders.

Calling of the function:

    negativeFolder = ('./negativeFolder');
    trainCascadeObjectDetector('trainedCrater1.xml',positiveData,negativeFolder);

The following is printed when the training starts:

    Automatically setting ObjectTrainingSize to [ 32, 32 ]
    Using at most 3992 of 4372 positive samples per stage
    Using at most 7984 negative samples per stage

    Training stage 1 of 20

The error that I am getting is this:

    Training stage 1 of 20
    [....................................................Error using ocvTrainCascade
    Error in generating samples for training. No samples could be generated for training the first cascade stage.

    Error in trainCascadeObjectDetector (line 265)
    ocvTrainCascade(filenameParams, trainerParams, cascadeParams, boostParams, ...

    Error in training (line 24)
    trainCascadeObjectDetector('trainedCrater1.xml',positiveData,negativeFolder);

Does anyone have any kind of suggestion on how to fix this issue? Thank you very much!

Answer 1

[Dima Lisin]

Hi Ferran,

Essentially, this error means that the first stage of your classifier has a very high false negative rate. In other words, it classifies most positive samples as negatives. It sounds like something may be wrong with your data. Can you please post some sample positive and negative images? What kind of objects are you trying to detect?

[Ferran Martinez]

Hello Dima, thank you for your answer!

I am trying to detect craters. This is a sample of the positive images: http://kdl.cs.umb.edu/w/wp-content/uploads/2012/05/examples3.jpg , and samples of negative instances can be seen here: http://pho.to/9TbAT

The dataset with both positive and negative images was published online as part of a research in which they used it for Haar-feature detection. I am not sure about which parameters of the training I could modify for it work.

Thanks again!

[Dima Lisin]

Hi Ferran,

This is a cool problem! But a hard one. There is a lot of variation in the images, especially caused by different amount of blurring. I would try using the HOG features, rather than Haar. Also, try selecting a small set of positive samples that are more homogeneous, and train on that. For example, pick only the samples that are in focus. You may have to split the "crater" class into multiple sub-classes, and train multiple detectors.

I would also try a smaller ObjectTrainingSize, such as 24x24. That may help reduce the effect of blurring.

[Ferran Martinez]

Hello Dima,

Thank you so much for all the suggestions! I'll try these variations.

[ronny3050]

Hi Ferran,

I was just wondering if you figured out a solution for the problem. I am stuck with the same situation. I have around 277 positive samples, however, all are very alike. My guess is that the classifier is failing because the positive samples are so alike that it cannot find addition features. I might be wrong, so I am wondering what you did to fix it.

Thank you so much. - Deb

[Eric]

Hi Ferran,

I was trying to train the cascade object detector in order to detect mouth region in the face. I was doing so because the built-in code that I've found on the internet was detecting eyebrow or nostrils as mouths in some of the test images. I have 30 positive images with ROIs and 10 negatives (without ROIs). When I started training the Viola-Jones classifier by trainCascadeObjectDetector function, I faced the same problem. Could you please tell me the solution? I hope you’ve found out the solution already.

Thank you -Eric.

Answer 2

[Zhenglin Wang]

Might be because training images are smaller than 32x32. Try to specify such as " 'ObjectTrainingSize', [28 28]"

[Hossam Alzomor]

what is the problem with small images?

I am facing the same problem, my images are 32*32

I have 106 ROI images and 96 NROI images

but the false -ve detection is 90%