Haarcascade classifier training

Question

I'm trying to create an xml file using samples I took with my camera. This is a test run where I put my camera against my window and let it take a picture every 30 seconds of passing cars for a while. I have now about 200 pictures (just for a small test), but I can't get any further.

I made a selecting tool to pick out the cars on pictures (bounding boxes) and mark pictures as negatives if there aren't any. Here are 2 examples of what the application looks like:

The application then saves the marked objects in positive.txt file as follows, skipping the non-marked pictures or the pictures marked as negative:

/samples/img_0000.jpg 1 74 66 34 30
/samples/img_0001.jpg 2 78 69 31 25 218 129 61 38
/samples/img_0003.jpg 1 83 72 21 21
/samples/img_0005.jpg 1 76 65 19 17
/samples/img_0006.jpg 1 127 112 37 24
/samples/img_0007.jpg 2 83 72 22 21 127 112 36 22
...

The negative pictures are simply saved to negative.txt file as follows:

/samples/img_0002.jpg
/samples/img_0004.jpg
/samples/img_0024.jpg
/samples/img_0026.jpg
...

Finally I try to run the haar training algorithm with /usr/bin/opencv_haartraining -data samples -vec positive.txt -bg negative.txt -npos 99 -nneg 20 -nstages 5 -mem 128 -minhitrate 0.999 -maxfalsealarm 0.5 -nonsym -mode ALL.

Don't mind the settings, I simply need this to work before I use a much more powerful computer to do the actual training on actual data with much more pictures.

For that I get the following output and error:

Data dir name: samples
Vec file name: positive.txt
BG  file name: negative.txt, is a vecfile: no
Num pos: 99
Num neg: 20
Num stages: 5
Num splits: 1 (stump as weak classifier)
Mem: 128 MB
Symmetric: FALSE
Min hit rate: 0.999000
Max false alarm rate: 0.500000
Weight trimming: 0.950000
Equal weights: FALSE
Mode: ALL
Width: 24
Height: 24
Applied boosting algorithm: GAB
Error (valid only for Discrete and Real AdaBoost): misclass
Max number of splits in tree cascade: 0
Min number of positive samples per cluster: 500
Required leaf false alarm rate: 0.03125

Tree Classifier
Stage
+---+
|  0|
+---+



Number of features used : 261600

Parent node: NULL

*** 1 cluster ***
OpenCV Error: Unspecified error (Vec file sample size mismatch) in icvGetHaarTrainingDataFromVec, file /build/opencv-XZa2gn/opencv-2.3.1/modules/haartraining/cvhaartraining.cpp, line 1930
terminate called after throwing an instance of 'cv::Exception'
  what():  /build/opencv-XZa2gn/opencv-2.3.1/modules/haartraining/cvhaartraining.cpp:1930: error: (-2) Vec file sample size mismatch in function icvGetHaarTrainingDataFromVec

Aborted

Does anybody know what that means? Why is there an error with sample size, whatever that is... I also tried replacing relative paths with absolute paths, but I got the same error. Is what I'm trying to do actually right, I haven't found any explicit examples on how to create a classifier from existing and marked pictures.

Answer 1

It turns out the .vec file is not simply the positive.txt file, it must be generated using opencv_createsamples.

So first I created the training_24-24.vec file (which is not just a text file) using the following command:

/usr/bin/opencv_createsamples -vec training_24-24.vec -info positive.txt -bg negative.txt -w 24 -h 24 -num 100

which created the training_24-24.vec file. Compared to my original question I remade the positive.txt file with square selections. Then I launched the classifier training with

/usr/bin/opencv_haartraining -data samples -vec training_24-24.vec -bg negative.txt -npos 99 -nneg 20 -nstages 5 -mem 128 -minhitrate 0.999 -maxfalsealarm 0.5 -nonsym -mode ALL

The performance is simply awful at this point, it recognises everything but the cars. This article says that one should use something like 1000 positive and 2000 negative images to start having something acceptable. Also it says for the -nstages option:

It’s useless to set many stage, if you have small number of positive, negative samples