Ask Your Question

Jaykob's profile - activity

2017-08-22 06:39:32 -0500 asked a question DNN performance on mobile platforms

Hi, I'm in the process of deciding whether to run my tensorflow model on iOS and Android using OpenCV's dnn module vs directly with tensorflow. Advantage using OpenCV would be that I don't need an additional lib (tensorflow) as I'm using OpenCV anyway. Disadvantage is that my model doesn't run out of the box due to some not yet supported layer types. But I think that could be solved somehow.

The main criteria would be performance, though. I only read that the dnn module got significantly faster with the 3.3.0 release, however I didn't find any comparison to tensorflow itself. Does anybody have some experience with both methods and could give me some hints?

2017-02-07 04:51:57 -0500 received badge  Supporter (source)
2017-01-31 02:59:47 -0500 received badge  Enthusiast
2017-01-30 04:42:55 -0500 commented question Structured edge detection performance

AFAIK OpenMP is not an option for iOS/Android, but correct me if I'm wrong. At least it's not building on my Mac right now. I might adopt the code to use TBB instead of OpenMP, ass TBB runs fine on both platforms.

2017-01-27 05:15:31 -0500 received badge  Nice Answer (source)
2017-01-27 05:10:01 -0500 commented question Structured edge detection performance

I didn't even know that I can influence this without modifying the code - or did I miss something? From looking briefly at the code I'm guessing 2 scales (1.0 and 0.5) are being used, though. Update: Images of size 400x300 take about 110ms on my machine

2017-01-27 05:05:27 -0500 received badge  Teacher (source)
2017-01-27 04:46:23 -0500 commented answer Training Neumann Matas classifiers

I know it's very late, but I finally uploaded the jTessBoxEditor to GitHub. Please see my note in the readme on how to invert the images as this is needed to train the erfilter.

2017-01-27 03:43:45 -0500 asked a question Structured edge detection performance

Hi, I'm using the StructuredEdgeDetection algorithm of the ximgproc module. Reading the according paper I was under the impression that it would be a fast algorithm suitable for realtime purposes.

However, it's taking 5.4 secs on my i7 MacBook Pro to process a 2592x1944 image. Is this resolution just too high or what could I do to improve the speed? I saw that parts of the code are supposed to make use of OpenMP. Since I'm targeting mobile devices (iOS, Android), I built OpenCV 3.2 with TBB support so far. Would you suggest modifying the code to make use of TBB instead of OpenMP in my situation or are there other things that I could try?

Thanks for your support!

2015-02-24 05:58:08 -0500 commented answer Training Neumann Matas classifiers

Hi, I sent you the code a while ago. Did you get it?

2015-02-18 09:05:33 -0500 received badge  Student (source)
2015-02-18 06:32:02 -0500 received badge  Editor (source)
2015-02-18 06:10:35 -0500 asked a question Android and OpenCV 3.0

I'm trying to build an app with native C++ code using OpenCV 3.0 and the Android NDK. As there is no Android SDK yet I'm wondering if there's already the possibility to build it from source? So far I managed to compile OpenCV 3.0 and build the android stuff under platforms/android. However I'm missing the Java classes now (BaseLoaderCallback for example).

Has anybody arrived doing this yet?

2015-01-30 05:52:55 -0500 answered a question Training Neumann Matas classifiers

You can use the following code to train your own classifiers: https://github.com/lluisgomez/erfilte... I just did it with success. For this process I modified the JTessBoxEditor application so that recognized characters can be cut out, inverted and exported as individual image files. This can save you much time by avoiding cutting out by hand. If you're interested I would upload it somewhere so that you can use it.

I'm currently trying to improve the grouping phase, because I also need to support single characters which are dismissed at the moment. Therefore I tried to train my own erGrouping classifier with the code on Lluis Gomez's GitHub account, but there are no sample images and my results aren't satisfying so far. So I got back to modifying the exhaustive search algorithm...

Update: I finally uploaded the modded jTessBoxEditor to GitHub. Please see my note in the readme on how to invert the images as this is needed to train the erfilter.