Segmentation fault if "char_width_mm = 20"

[Silex]

Hi,

train-ocr wouldn't recognize the I in http://www.licenseplates.tv/images/swissai.gif, so I reduced char_width_mm to 20 for the [eu] country and ./classifychars now display the I correctly, however when I try to press enter to input which char it is then it segfaults.

[matthill]

I think the setting you want to modify is:

segmentation_min_box_width_px = 8

That's a value in pixels for the minimum possible character width, anything less gets disqualified. Setting it to 4 should pass the "I" through.

I recropped the image before processing it. I assume you're doing this as well. But my crop may have been slightly different, so it may not have hit the exact same logic to reproduce the crash.

[Silex]

I confirm that setting it to 4 prevents the segfault. Thanks!

[Silex]

Side question, for train-detector I should give "real-life" images, does the same holds for train-ocr?

Basically my questions are:

• Should I give very good images to classifychars or it doens't matter?
• it looks like adding the swiss plate ocr data to the eu dataset doesn't yield as good results as creating a entirely new country for switzerland... basically, lots of A from eu dataset are mixed with 4 from the swiss plate...
  • Can we tell a preference order when creating leu.traineddata?
  • If I want accuracy, should I consider splitting different datasets into different countries and call ./alpr each time with a different country and pick the one with the highest %?

[matthill]

You should give the detector the exact type of images that you want it to detect. I usually include a little bit of car bumper/edges -- just an inch or two -- to give it some context (the ALPR stages will filter this stuff out later). For slightly rotated plates, this may not even include the full plate (e.g., a plate rotated counter-clockwise may be missing the a little bit of the bottom left and top right portions).

Here's a video demonstration of how Cascade classification works -- seeing something like this made it much more clear: https://www.youtube.com/watch?v=nVbaNcRldmw

OCR is a little different. You want to give it actual images from actual plates, but you don't want to give it characters that are missing sections or "damaged" in some way. You want the characters to look relatively clean. Tesseract needs a good sample, and it can work out its best guess when actual char images are given and are missing portions.

[Silex]

Alright, so :

• train-detector: real data, even of bad quality
• train-ocr: real or fake data, but always good quality (by fake I mean http://tinyurl.com/nzudrdx which is not a "real" plate.)