1

Topic: Application of detectors of objects

Advise what to esteem, please, on any subtleties \cunnings of application of detectors of objects on images, from the point of view of model usage as black box. Model training, unconditionally, difficult, but besides it it is necessary to create an infrastructure of its application and post-handling of its results. For example, there is a qualifier which is caused in a sliding window and can work for one object in two adjacent windows. If I have to count finally an amount of objects it is necessary to unite somehow these adjacent actuatings, and here it is possible to offer different methods. But it is similar to a widespread problem and for certain for it there are standard decisions. The question was not specific about this task, and is faster about, whether there are sites \blogs \books etc., to esteem about the stored practical experience of usage ml (especially with images) with all accompanying complexities?

2

Re: Application of detectors of objects

Hello, tdiff, you wrote: T> the Question was not specific about this task, and is faster about, whether there are sites \blogs \books etc., to esteem about the stored practical experience of usage ml (especially with images) with all accompanying complexities? Here a small dial-up of transfers and reviews written by accessible language: http://mechanoid.kiev.ua

3

Re: Application of detectors of objects

Hello, tdiff, you wrote: T> For example, there is a qualifier which is caused in a sliding window and can work for one object in two adjacent windows. If I have to count finally an amount of objects it is necessary to unite somehow these adjacent actuatings, and here it is possible to offer different methods. But it is similar to a widespread problem and for certain for it there are standard decisions. Yes, suppression  or Non-Maximum-Suppression is called. T> the Question was not specific about this task, and is faster about, whether there are sites \blogs \books etc., to esteem about the stored practical experience of usage ml (especially with images) with all accompanying complexities? It seems to me that it is better to be repelled from the specific target decision. Generally, skilled people say that  here out of competition and in the foreseeable future the competitor it is not visible. Classical algorithms are sense to select, more likely, for speed and on well studied problems. But also so not always.

4

Re: Application of detectors of objects

Hello, Nik, you wrote: N> Here a small dial-up of transfers and reviews written by accessible language: N> http://mechanoid.kiev.ua Thanks, I will remember the link.

5

Re: Application of detectors of objects

Hello, Nuzhny, you wrote: N> Yes, suppression  or Non-Maximum-Suppression is called. "algorinth"

6

Re: Application of detectors of objects

Hello, Nuzhny, you wrote: N> Yes, suppression  or Non-Maximum-Suppression is called. Thanks N> seem To me that it is better to be repelled from the specific target decision. Generally, skilled people say that  here out of competition and in the foreseeable future the competitor it is not visible. Classical algorithms are sense to select, more likely, for speed and on well studied problems. But also so not always. And about  it to what? Here these "infrastructural" things - for networks in another way look?

7

Re: Application of detectors of objects

Hello, tdiff, you wrote: T> And about  it to what? Here these "infrastructural" things - for networks in another way look? In many respects - yes. And it is not so much because they are not necessary, how many because they are already included in more high-level decisions and most to apply them specially it is not necessary. To take the same suppression  from an example above. Decided to use YOLO - all already there is. To use same of OpenCV - too is. You use any keras - all is. If most to take HOG+SVM or Haar+Boost  it will be already necessary to care of suppression: most to write, cause ready or to specify parameter in function.

8

Re: Application of detectors of objects

Hello, Nuzhny, you wrote: N> In many respects - yes. And it is not so much because they are not necessary, how many because they are already included in more high-level decisions and most to apply them specially it is not necessary. It turns out what to use more difficult tools in something even easier, than simple

9

Re: Application of detectors of objects

Hello, tdiff, you wrote: T> It turns out what to use more difficult tools in something even easier, than simple Yes. The Same enough simple and clear SVM in practice to apply difficult enough. To select constants, features, it is sensitive very much to noise etc. But in the linear variant is fast, good and clear.