Is it possible to prepare data with opencv_createsamples for later cascade training?

Question

Image processing

I have been suffering for a long time, trying to teach the classifier to recognize any objects, on the data that I generated from one image. With manual designation, everything worked out, but with automatic designation, the classifier stops teaching at the second or third level, and the results of searching for objects by this classifier (with 2-3 levels) leave much to be desired. The desired objects are usually simple: company logos, characteristic icons.

Answer 1

1. If we are talking about a cascade, then they are more likely not for these tasks, for yours (logo, characteristic icon) SURF habrahabr.ru/post/155651 is much more suitable
2. Training may stop due to a poor negative / positive ratio, also you need to keep in mind that the sample size should be significant (we trained on 5000 positives and 7000 negatives), with a significant increase in negatives, the training lasted a very long time and the quality was poor, usually, if everything goes well, then training rarely takes more than two hours (well , specifically in our case). Therefore, it makes sense to play around with the number of positives and negatives. This article helped at one time: note.sonots.com/SciSoftware/haartraining.html