Let’s create a detector.

I will train the classifier with training windows of size 50 x 42 :

WIDTH=50
RATIO=0.85

where ratio is height divided by width.

The dimensions specify the smallest object size the classifier will be able to detect. Objects larger than that will be detected by the multiscale image pyramid approach.

Extracting rectangles to OpenCV format

As a best practice, I would recommend to create an executable, extract, to extract training windows, positive ones as well as negative ones, from an annotated input of your choice :

./extract input.csv $WIDTH $(expr $WIDTH*$RATIO |bc)

The purpose of my extract program is to create two directories that can be directly used by OpenCV cascade training algorithm :

pos
-- info.dat
-- img
---- xxx.png
---- yyyy.png
neg
-- info.dat
-- img
---- zzzz.png
---- llll.png

neg/info.dat is a simple list of images

img/zzzz.png
img/llll.png

while pos/info.dat contains rectangle informations

img/xxxx.png 1 x y w h
img/yyyy.png 1 x y w h

In the pos/img/ images are full size, since the rectangle information is in the info.dat file, whereas the neg/img/ images are extracted.

In my case I provide many more negatives than positives to the classifier (4 times more).

I would avoid leaving OpenCV training algorithm create all the negative windows (opencv_traincascade subsample negative image), or to do that, my extract will create the background images at the final training size (100x20 in my example) so that it cannot subsample but only take the entire negative image as a negative.

Creating negatives from the backgrounds of the positives is much more “natural” and will give far better results, than using a wild list of background images taken from the Internet. That’s all that makes the interest of such an extract program.

I will add these negatives to negatives from this repo :

cd ~/apps
git clone https://github.com/christopher5106/tutorial-haartraining.git
cd data/negatives
ls -l1 *.jpg > negatives.txt

be careful: in some cases, OpenCV requires an absolute path for images in the negatives.txt (otherwise you could get an error Train dataset for temp stage can not be filled. Branch training terminated.)

The CSV input file to the program is a list of input images with the class and coordinates of the rectangles where objects are located in the image,

/Users/christopherbourez/data/img.png,0,10,30,210,65

The last two input parameters give the size to resize the negative windows after extraction.

OpenCV positives preprocessing

Let’s set the number of positives we take (NUMPOS) :

NUMPOS=1000

It is required to use an OpenCV program to convert the positive rectangles to a new required format :

cd MYPOSITIVES_FOLDER
opencv_createsamples -info pos/info.dat -vec pos.vec -w $WIDTH -h $(expr $WIDTH*$RATIO/1 |bc) -num $NUMPOS

You could also augment the positive sample by rotating and distorting the images with opencv_createsamples and merging them back into one vec with Naotoshi Seo’s mergevec.cpp tool.

Train the classifier

Let’s the number of negatives we take per positives (FACTOR) :

FACTOR=10

and launch the training :

mkdir data

opencv_traincascade -data data -vec pos.vec -bg ~/apps/tutorial-haartraining/data/negatives/negatives.txt -w $WIDTH -h $(expr $WIDTH*$RATIO/1 |bc) -numPos $(expr $NUMPOS*0.85/1 |bc) -numNeg $(expr $FACTOR*$NUMPOS*0.85/1 |bc)  -precalcValBufSize 1024 -precalcIdxBufSize 1024 -featureType HAAR

About the training parameters :

  • I fix numPos parameter to be about 90% of the number of positive rectangles, since some positives that are too different from the the positive set can be rejected by the algorithm and if numPos equals the number of positives, it will fail with the following message :

      OpenCV Error: Bad argument (Can not get new positive sample. The most possible reason is insufficient count of samples in given vec-file.

    Increasing the number of positives will enable a better generalization of the model. Usually a few thousand is good.

  • numNeg : it is usually good to take two times more negatives than positives.

    Increasing the number of negative will diminish the number of false positive detections.

  • increasing numStages will not improve anymore the model when overfitting occurs. In this case, you’ll need to add more positives and negatives to the training set.

Be careful also, the JS library jsfeat only accept detectors in the old format (use opencv_haartraining instead).

Use in nodeJS

Simply create a recognize.js program :

var cv = require("opencv");

var color = [0, 255, 0];
var thickness = 2;

var cascadeFile = "models/cascade.xml";

var inputFiles = [ "image.jpg" ];

inputFiles.forEach(function(fileName) {
  cv.readImage(fileName, function(err, im) {
    im.detectObject(cascadeFile, {neighbors: 2, scale: 2}, function(err, objects) {
      console.log(objects);
      for(var k = 0; k < objects.length; k++) {
        var object = objects[k];
        im.rectangle([object.x, object.y], [object.width, object.height], color, 2);
      }
      im.save(fileName.replace(/.jpg/, "processed.jpg"));
    });
  });
});

and call the detector

node recognize.js

Well done !

more

A few posts : 1

  • 2
  • 3

  • To have a quick start, try this example

    git clone https://github.com/mrnugget/opencv-haar-classifier-training.git