opencv_traincascade : Acceptance Ratio- Unconsistent Behavior

asked 2015-05-03 08:32:10 -0500

pulp_fiction gravatar image

updated 2015-05-05 08:02:05 -0500

I was training a classifier with opencv_traincascade. I trained till 3 stages then stopped. I resumed the training after some time when I came back, I saw that when restarting the training, the acceptance ratio has changed(decreased) compared to what I saw in the half done training of 4th stage when I had to stop it.

Anyone knows, what is the issue here? Is it worth bothering?

edit retag flag offensive close merge delete


You can only see a correct acceptance rate estimation after it finishes your training stage and calculates the influence on the training data for the new stage. I have done several retraining of models with exactly the same parameters and the models are 100% identical in final values. You sure you didn't change a thing?

StevenPuttemans gravatar imageStevenPuttemans ( 2015-05-04 03:42:42 -0500 )edit

@StevenPuttemans : Nope, didn't change a thing. I have had this situation a few times before!

pulp_fiction gravatar imagepulp_fiction ( 2015-05-04 04:33:13 -0500 )edit

Hmmm so if you train a model of 3 stages, and then restart with the same data a model of 3 stages, then you have different values? Could you try that?

StevenPuttemans gravatar imageStevenPuttemans ( 2015-05-04 06:02:45 -0500 )edit

@StevenPuttemans I personally learned a lot by reading your comments here and there, you are right my comment may not help him even a bit but Ive just shared that I had this issue before (with fine working final result) so a pro like you could make an estimation of how reproducible is this issue on different environments.

theOneCV gravatar imagetheOneCV ( 2015-05-05 00:59:37 -0500 )edit

@theOneCV : You are not wrong entirely. In my previous training I had the training disrupted and the ratio went down. I ignored it, tested the classifier, it worked as expected (actually because I had trained it for only 10 stages, it showed expected progress). I asked the question this time because of the time it takes. I am skeptical because I am a beginner.

pulp_fiction gravatar imagepulp_fiction ( 2015-05-05 06:01:15 -0500 )edit

@StevenPuttemans : I wanted to ask a thing : I have been training with -w 60 and -h 50 uptil now. In Naotoshi Seo's notes on the training he has trained the -w 20 and -h 20. My question is : Should I decrease the current width and height ratio furthermore? Is it gonna result in better classifier?

pulp_fiction gravatar imagepulp_fiction ( 2015-05-05 07:24:39 -0500 )edit

It depends on what you want. If you want to detect objects that are smaller than 60x50 pixels, than yes reduce these dimensions, if not, than keep them because they will yield more features (more detail) and thus possible a better classifier.

StevenPuttemans gravatar imageStevenPuttemans ( 2015-05-07 05:02:32 -0500 )edit