Phinatic's profile - activity

Hello,

Does the -mode parameter affect training with local binary pattern features? From my understanding it won't, since -mode uses either basic upright haar features or more advanced ones proposed by Lienhart and Maydt (from my understanding). I think LBP is a different set of features altogether? I have also found no evidence of it in the sourcecode.

I am just asking this question to be sure about my assumption, since I could have easily overlooked something :)

Hello all,

I am currently using the opencv_traincascade application to train a cascade classifier. I have read that it is possible to pause the program (and even use the intermediate result). How should I go about doing this? Can I just kill the program (ctrl+c)? And how do I continue where I left off later?

I just want to make sure I don't lose all of my data- it has been training for well over a day so far.

Thanks in advance.

I found both pages on my own too- I'm not sure how I missed line 360. Will keep in mind that the first ones are usually defaults, Thank you for all the support and have a nice day!

Thank you.

Thank you. I'm not sure how I did not stumble upon it reading documentation though. Could you give me a reference to where I could find that the default specifies N=9?

EDIT: This question has been answered. I have tried to close this, but my lack of reputation does not allow it. I welcome anyone with the permission to close this to do so.

Hi,

I wanted to examine which points in the image were picked up by using the default FAST Detector. I currently use EmguCV to do so, though I think my settings are the same as in this guide (I am calling the default constructor, tried different threshold values and non-max suppression on). I went through the wikipedia article and the guide I pointed to earlier, and tested it on one part of my image. This was the result .

From my understanding, FAST determines whether or not there is a corner based on the following circle of grayscale values (taken from the image, but for clarity):

It is a corner when at least N contiguous pixels are either threshold T brighter or darker than the center point. In this case I think N=12 (as default), and (at first) picked T = 10. This means that this segment of the picture should fail to mark as a corner, since the middle-top and left-middle pixels both have a value of 219, but the middle-bottom and right-middle pixel have 41.

Varying the threshold T however, I did find that this spot does not get detected when T >= 85. This is probably because the top-right and bottom-left pixels have an intensity of 161 (76+85). Could anyone help me explain this behavior? Why does it find the spot at T < 85? Is the default N < 12? Or am I missing some step here?

Thanks in advance.