About opencv_traincascade....

edit

I read it and I agree ;-) Very helpful indeed!

@Mathieu Barnachon thank you!

I've started reading the book, but in the meantime I'd like to ask a few things to leave it training while I read. I read that you answered that it's better to have 50 good positive images than getting one and generate 50 with opencv_createsamples. Therefore, as I can take 50 images of my object from different angles and use them as positives. Would this be better? The other thing is the negatives. As I want to detect the objects in a controlled environment (e.g my office) I can do a 'random' walk gathering images without the object, right?

I also read that I should aim for NEG count : acceptanceRatio around 0.0004 to consider a good cascade and if it is ~5.3557e-05 over trained?

Forgot to ask. By default the boosting algorithm is GAB. Which one gives better results? I believe the chose GAB as it is the one that consumes less ram? The computer I'm using has 32gb of ram plus another 32gb of swap so, should I go with RAB?

Look at the updates in my answer ;) will add the last question also!

I've just gathered 60 images from the phone different angles (a few degrees...) and hights with 300 negative samples of randomly walking in the office....training took 1 o 2 minutes, got a Required leaf false alarm rate achieved with NEG count : acceptanceRatio 250 : 0.000752328 but still got around 25 false positives......I trained it with RAB....

@aripod detecting object completely rotation and viewpoint invariant is indeed still something that can NOT been done in real time. The techniques and hardware needed for that are simply not yet standard, not even in research communities. It differs if you have a set of known parameters like symmetry, camera distance, lighting conditions, ... but then it is still very challenging! As to the turning table, it won't work since the object will have the background of the rotation table, and thus the edge features will not be representative for decent use, leading to false positive detections or no detections at all. Precision and recall calculations, take a look at wikipedia, it explains it pretty well.

For the model trained, there is a difference between negative images and negative samples. Can you give me your training parameters? Then it will become clearer what you did. Your model actually tells you that it cannot remove negative windows as negative, so you need more negatives. Are your actual objects being detected? If not, you need more positives.

Maybe I don't use a rotating table but "I" rotate around the object to get the background? Here is my dataset. The object is rarely detected. Once again, thanks for your big help!

Not sure why that would not work, it quite the distinctive object you are using ....

I don't believe there's something wrong with the line to run the training" opencv_traincascade -data data -vec positives.vec -bg bg.txt -numPos 150 -numNeg 300 -numStages 10 -w 24 -h 24 ....

Actually there is. You have 300 negative images, but are passing 300 to -numNeg parameter which is actually the number of negative windows which is quite different to the amount of images. Also only 300 negatives to 200 positives will never yield any good classifiers. Try at least like 250 POS and 1000 NEG windows. Basically you have not enough training data for the moment to yield a better classifier.

Sorry for the dumb question..A window would be a section of an images, right?

Well the model is trained by resizing your positive annotations to the -w and -h parameters first. Then a sliding window of -w x -h size goes over negative images and cuts out negative samples. BTW why are you making those parameters only 24 pixels large?

Ok, got it. So I don't need to get more neg but increase the windows for the training algorithm, but for the 250 pos I would need around 280 positives, right? I chose 24 based on the frontal_face cascade.....plus if I make it bigger, the detection would be slower? I don't mind if the training is slow, but detection should be fast as there would be more than 5 objects to detect, and considering a few cascades per image to have it rotation variant, it might take a long time to run de detection (it needs to be ~ real time)

I took 250 new images of the object (with different backgrounds) and 305 images of the office without the object and left it training using opencv_traincascade -data data -vec positives.vec -bg bg.txt -numPos 220 -numNeg 1000 -numStages 10 -w 80 -h 80 -bt RAB. After 10 and a half hours I got:

POS count : consumed   220 : 221
NEG count : acceptanceRatio    1000 : 0.000370403
Required leaf false alarm rate achieved. Branch training terminated.

The result of the detection is: without the object there are false detections and with the object there are also false detections plus no good ones....

And how are you setting your detection parameters? You must be doing something wrong here ... because your result seems impossible. Can you give me your detection code? Also could you change the boosting the standard for a starter?

I found the issue! In the detection I'm downscaling so it is faster and I commented the rescaling of the bounding box. The code is in the question, as EDIT 1. Now the detection works but I still have a lot of false positives. Look in this short video

Ow your code detect_cascade.detectMultiScale(frame_gray, objects, 1.2, 2, 0|CV_HAAR_SCALE_IMAGE, Size(80, 80)); explains it all. You are placing the number of adjacent neighbors on 2 which means each weak detection is visualised also. Stuff you can do to improve this

• Increase the minNeighbours parameter to at least 5 or 10. This will ensure that only higher scoring detections remain.
• However still some FP detections can trigger a high certainty. Why not use the temporal approach and smooth over frames. Only if a detection remains for more then 4-5 frames, you visualise it, else you remove it. This will also remove alot of FP detections.

I changed minNeighbours to 5 and then 10 but I still have a lot of false positives. For the temporal approach, these false detections also remain in time so they would still be detected and never discarded..... I'm running the training again by setting the maxFalseRateAlarm to 0.25 (half default), I switch back to GAB. And in parallel I'm running the same but setting to LBP rather than HAAR to see which one works better......what do you think?

Should I get more positive samples to avoid the false negatives....? Or increase the negative windows?

I would go for more negatives, since it can clearly find your positives. Try increasing it to lets say 2000 samples per stage. Also LBP trains alot faster than HAAR, stick to that for the time being. Putting maxFalseAlarmRate to 0.25 seems wrong to me, since the doesnt do what we expect from cascade classifiers. You will train harsher singles stages and thus a lot more features will need to be evaluated. You want to discard as many negatives as possible with the least effort as possible.

Should I get more negative images of with the ones I have it's enough? Just increasing the numNeg will be enough? And you think I should do all this tests to LBP until getting the right training values and then retrain it with HAAR? And what about the rotation of the object? It keeps detecting it when I rotate close to 45 deg to each side...

You will never now if you have enough negative samples to grab windows from as long as it doesn't throw an error. So start by simply increasing numNeg value. Yes, do LBP untill you get decent results, the drawback in accuracy is so minimal that I am not even using HAAR features anymore. Depending on the model and viewpoint it is indeed rotation invariant on itself. I guess the border is around +-30 degrees. But you won't cover the full 360 degrees with one model.

And I guess the bigger the numNeg the better, more robust, right? I'll keep testing with that number with the negative samples I have for now and get back to you with the results. Is there a way to specify how it gets the negative samples? It might be nice to specify a number of windows for each image to have this process more "controlled" and gather the full picture rather than randomly For the rotation, I don't need 360, but around 45 degrees to each side.....which for now seems correct.....in the case I want to increase this, the solution is ALWAYS to get more samples, right?