Windows vs Linux installation

[Hannes kruger]

I used the pre compiled windows program and got a perfect score with my sample license plates. alpr 1.2.0.

With linux I compiled openalpr 1.2.0 with opencv 2.4.10 , tesseract 3.03 and  leptonica-1.71.

With linux I got very bad results. Not even 50%. I compared the config files and trained data. It is the same for both operating systems.


Where can I look find out why the linux run is so bad. Linux is the platform I need to use.

[Matt]

The same code will give you the exact same results regardless of platform.  Do you see any warnings or anything in the output?  Any compiler errors or warnings?

[Piotr.23]

Hi, am also getting nice results with the Windows version and bad with Linux version. 

I go the following error codes during the "Easy installation" following https://github.com/openalpr/openalpr/wiki/Compilation-instructions-(Ubuntu-Linux)

[ 45%] Built target alpr

Scanning dependencies of target alprd

[ 46%] Building CXX object CMakeFiles/alprd.dir/daemon.cpp.o

[ 48%] Building C object CMakeFiles/alprd.dir/daemon/beanstalk.c.o

/home/piotr/_download/openalpr/src/daemon/beanstalk.c: In function ‘bs_put’:

/home/piotr/_download/openalpr/src/daemon/beanstalk.c:380:5: warning: format ‘%lu’ expects argument of type ‘long unsigned int’, but argument 7 has type ‘size_t’ [-Wformat=]

     snprintf(command, 1024, "put %"PRIu32" %"PRIu32" %"PRIu32" %lu\r\n", priority, delay, ttr, bytes);

     ^

/home/piotr/_download/openalpr/src/daemon/beanstalk.c:380:5: warning: format ‘%lu’ expects argument of type ‘long unsigned int’, but argument 7 has type ‘size_t’ [-Wformat=]

/home/piotr/_download/openalpr/src/daemon/beanstalk.c: In function ‘bs_reserve_job’:

/home/piotr/_download/openalpr/src/daemon/beanstalk.c:450:9: warning: format ‘%lu’ expects argument of type ‘long unsigned int *’, but argument 4 has type ‘size_t *’ [-Wformat=]

         sscanf(message->status + strlen(bs_resp_reserved) + 1, "%"PRId64" %lu", &job->id, &job->size);

         ^

/home/piotr/_download/openalpr/src/daemon/beanstalk.c: In function ‘bs_peek_job’:

/home/piotr/_download/openalpr/src/daemon/beanstalk.c:534:9: warning: format ‘%lu’ expects argument of type ‘long unsigned int *’, but argument 4 has type ‘size_t *’ [-Wformat=]

         sscanf(message->status + strlen(bs_resp_found) + 1, "%"PRId64" %lu", &job->id, &job->size);

         ^

/home/piotr/_download/openalpr/src/daemon/beanstalk.c: In function ‘bs_get_info’:

/home/piotr/_download/openalpr/src/daemon/beanstalk.c:599:9: warning: format ‘%lu’ expects argument of type ‘long unsigned int *’, but argument 3 has type ‘size_t *’ [-Wformat=]

         sscanf(message->status + strlen(bs_resp_ok) + 1, "%lu", &size);

         ^

[ 49%] Building CXX object CMakeFiles/alprd.dir/daemon/beanstalk.cc.o

Linking CXX executable alprd

Can you help, please? | P

[Vladimir Radnovic]

ALPRD - deamon don't work on windows

u can cansell intalations and you will get all work

[Piotr.23]

I have no problems with Windows. My Ubuntu version doesn't work as expected (with EU plates; US plates appear to work) | P

[Vladimir Radnovic]

What you use ? alpr -c eu nameimage.jpg or alprd ? 

can u past here configure file 

[Piotr.23]

I use  alpr -c eu nameimage.jpg

[Piotr.23]

Do you mean openalpr.conf.in?

[common]

; Specify the path to the runtime data directory

runtime_dir = ${CMAKE_INSTALL_PREFIX}/share/openalpr/runtime_data

ocr_img_size_percent = 1.33333333

state_id_img_size_percent = 2.0

; detection will ignore plates that are too large.  This is a good efficiency technique to use if the 

; plates are going to be a fixed distance away from the camera (e.g., you will never see plates that fill 

; up the entire image

max_plate_width_percent = 100

max_plate_height_percent = 100

; detection_iteration_increase is the percentage that the LBP frame increases each iteration.

; It must be greater than 1.0.  A value of 1.01 means increase by 1%, 1.10 increases it by 10% each time. 

; So a 1% increase would be ~10x slower than 10% to process, but it has a higher chance of landing 

; directly on the plate and getting a strong detection

detection_iteration_increase = 1.1

; The minimum detection strength determines how sure the detection algorithm must be before signaling that 

; a plate region exists.  Technically this corresponds to LBP nearest neighbors (e.g., how many detections 

; are clustered around the same area).  For example, 2 = very lenient, 9 = very strict.

detection_strictness = 3

; The detection doesn't necessarily need an extremely high resolution image in order to detect plates

; Using a smaller input image should still find the plates and will do it faster

; Tweaking the max_detection_input values will resize the input image if it is larger than these sizes 

; max_detection_input_width/height are specified in pixels

max_detection_input_width = 1280

max_detection_input_height = 720

; Bypasses plate detection.  If this is set to 1, the library assumes that each region provided is a likely plate area.

skip_detection = 0

max_plate_angle_degrees = 15

ocr_min_font_point = 6

; Minimum OCR confidence percent to consider.

postprocess_min_confidence = 65

; Any OCR character lower than this will also add an equally likely

; chance that the character is incorrect and will be skipped.  Value is a confidence percent

postprocess_confidence_skip_level = 80

; Reduces the total permutations to consider for scoring.

postprocess_max_substitutions = 2

; Results with fewer characters will be discarded

postprocess_min_characters = 4

postprocess_max_characters = 8

[debug]

general       = 0

timing        = 0

state_id      = 0

plate_lines   = 0

plate_corners = 0

char_segment  = 0

char_analysis = 0

color_filter  = 0

ocr           = 0

postprocess   = 0

show_images   = 0

pause_on_frame = 0

;;; Country Specific variables ;;;;

[us]

; 30-50, 40-60, 50-70, 60-80

char_analysis_min_pct = 0.30

char_analysis_height_range = 0.20

char_analysis_height_step_size = 0.10

char_analysis_height_num_steps = 4

segmentation_min_box_width_px = 4

segmentation_min_charheight_percent = 0.5;

segmentation_max_segment_width_percent_vs_average = 1.35;

plate_width_mm = 304.8

plate_height_mm = 152.4

multiline = 0

char_height_mm = 70

char_width_mm = 35

char_whitespace_top_mm = 38

char_whitespace_bot_mm = 38

template_max_width_px = 120

template_max_height_px = 60

; Higher sensitivity means less lines

plateline_sensitivity_vertical = 25

plateline_sensitivity_horizontal = 45

; Regions smaller than this will be disqualified

min_plate_size_width_px = 70

min_plate_size_height_px = 35

ocr_language = lus

[eu]

; One-line European style plates

; 35-50; 45-60, 55-70, 65-80, 75-90

char_analysis_min_pct = 0.35

char_analysis_height_range = 0.15

char_analysis_height_step_size = 0.10

char_analysis_height_num_steps = 5

segmentation_min_box_width_px = 5

segmentation_min_charheight_percent = 0.4;

segmentation_max_segment_width_percent_vs_average = 2.0;

plate_width_mm = 520

plate_height_mm = 110

multiline = 0

char_height_mm = 80

char_width_mm = 53

char_whitespace_top_mm = 10

char_whitespace_bot_mm = 10

template_max_width_px = 184

template_max_height_px = 46

; Higher sensitivity means less lines

plateline_sensitivity_vertical = 18

plateline_sensitivity_horizontal = 55

; Regions smaller than this will be disqualified

min_plate_size_width_px = 100

min_plate_size_height_px = 20

ocr_language = leu

[Vladimir Radnovic]

plz add one image here to test it

tnx

vladimir

[Piotr.23]

Thanks, Vladimir. 

I am attaching two images: 12 and 13. 

12 was perfectly recognized, and 13 almost perfectly recognized on Windows installation.
12 and 13 were initially not recognized on Ubuntu

piotr@piotr:~$ alpr -c eu --clock -n 1 /home/piotr/_download/3-20150208/3-20150208-0113-ALPR.jpg

Total Time to process image: 96.2574ms.

No license plates found.

I then tweaked alpr config file and changed:

- detection_iteration_increase = from 1.1 to 1.5

- detection_strictness = from 3 to 6

- max_plate_angle_degrees = from 15 to 30

Following this, image 12 was recognized and 13 still not. 

This is just an example, I tested many images. 

Windows is always much better than Ubuntu version

And I need Linux to have all on one system

Yes, there are some extra characters on the images, but they never confuse ALPR

Best | P

[Vladimir Radnovic]

For start tray to put camera litle left, if you can put better resolution now is very small 768x512 optimal will be 1280x720

on second image plate is not good it's par of image/plate...

second remove txt from image ( on camera ) it's better

3) better result you will have if we read plate from back of care... in night time... you don't have light on front

In attach you have some exp image how to set up camera...

[Piotr.23]

Vladimir: the conditions are not perfect, but can't change them

However, this still does not explain the difference between Win and Linux

Here are the results of your examples:

F:\$Download\openalpr\ffmpeg\test>F:\$Download\openalpr\alpr.exe -c eu -n 1 test.jpg

plate0: 1 results

    - LSA001NN   confidence: 90.1882     template_match: 0

F:\$Download\openalpr\ffmpeg\test>F:\$Download\openalpr\alpr.exe -c eu -n 1 test2.jpg

plate0: 1 results

    - KI002YG    confidence: 88.7756     template_match: 0

piotr@piotr:~$ alpr -c eu --clock -n 1 /home/piotr/_download/test/test.jpg

Total Time to process image: 456.583ms.

No license plates found.

piotr@piotr:~$ alpr -c eu --clock -n 1 /home/piotr/_download/test/test2.jpg

Total Time to process image: 297.589ms.

plate0: 1 results -- Processing Time = 130.814ms.

    - KI002Y     confidence: 89.9517

Windows found partially correct number on 'test' and correct on 'test2'

Ubuntu did not find any on 'test' and partially correct on 'test2'

How to change Linux conf to match Win version? | P

[Vladimir Radnovic]

it's 2 difrent version od alpr windows and linux

tipe alpr --version on linux and win. and past here

tnx

vladimir

[Matt]

Are the Windows and Linux software running the exact same software code?  Did you compile both yourself?  What do you get when you type alpr --version on Windows and on Linux?

Are the runtime_data/region/eu.xml files the same?  Perhaps run a diff against them to see.

Are the openalpr.conf files the same?  Perhaps run a diff against them as well.

[Piotr.23]

Apologies: I badly formatted the text of Windows files for diff comparison. 

NO: there are no differences in eu.xml file.

YES: there are small differences in openalpr.config file:

4c4

< runtime_dir = /usr/share/openalpr/runtime_data

---

> runtime_dir = C:/set/to/runtime_dir/path

20,21c20

< ; Piotr 20150217 was 1.1

< detection_iteration_increase = 1.5

---

> detection_iteration_increase = 1.1

26,27c25

< ; Piotr was 3

< detection_strictness = 6

---

> detection_strictness = 3

36,37c34

< ; Bypasses plate detection.  If this is set to 1, the library assumes that each region provided is a likely plate area.

< skip_detection = 0

---

> multithreading_cores = 1

39,40c36,38

< ; Piotr was 15

< max_plate_angle_degrees = 30

---

> 

> 

> max_plate_angle_degrees = 15

60c58

< timing        = 1

---

> timing        = 0

63a62

> char_regions  = 0

89,90d87

< multiline = 0

< 

110d106

< ; One-line European style plates

124,125d119

< 

< multiline = 0

P

On Wednesday, February 18, 2015 at 9:21:16 PM UTC+1, Piotr.23 wrote:

Hi, Matt

I did not compile Windows version. I just downloaded the exe file, its version is 1.2.0. I even did not realize that there is something to compile, it works OK on Windows 7

The way I downloaded Linux version (2.0.0) I described earlier in this thread. I used "Easy installation" following https://github.com/openalpr/openalpr/wiki/Compilation-instructions-(Ubuntu-Linux)

There is a number of differences in both eu.xml and openalpr.conf. I am attaching these files here.

What shall I change in Linux configuration to have as good results as with Win 1.2 version? | P

[Matt]

It looks like you're comparing the 1.2 version compiled on Windows with the 2.0 version compiled on Linux.  The results will be different, since the code is different.

There's more discussion on this here: https://groups.google.com/forum/#!searchin/openalpr/1.2/openalpr/vpky-7qFXSI/lD9muNs6n6wJ

Whenever I make changes to the library, I run the new code against a benchmark test of many images.  There will be variations between versions, but the general trend (ideally) is increased accuracy and performance.

-Matt

[Piotr.23]

Matt, I forgot to thank you for developing this marvelous project :)

Yes, I am comparing these versions, although, these are not apples and oranges... 

For the reason I'd like to identify, the older Win version appears to work better than recent Linux version, both on my samples as well as on samples provided by Vladimir. 
Could this be due to 32 (Linux)/64 (Windows) - bit systems? | P

[Matt]

Much has changed between version 1.2 and 2.0, there's been almost 250 commits since the last release.  

I don't think we can claim that the accuracy has gone down if you're just comparing a handful of sample images.  We must also consider all the images that were not recognized in the past version and now are recognized.  With computer vision, even if you change a few pixels in an image, or slightly tweak the weighting rules/scores, the results can be completely different.  So there is an element of randomness to it.  

I have a large library of benchmark images that I run against the software with each new version.  At the end of the run I get an accuracy score.  The numbers between 1.2 and 2.0-dev have gone up in my benchmarks for both EU and US.  So, I just think the differences you are seeing are due to a small sample set.  Also, I don't believe that there is any difference in results between Windows and Linux when they're running the same version of the software.  Nor is there any difference between 32 and 64 bit.

[Piotr.23]

Matt, you are absolutely right: only the comparison of a large set of images makes sense. 

Please forgive my ignorance... how to use alpr on several images? 

I have original AVI files. Since Windows version does not allow me to assess AVIs (Linux does)... in WIndows, I first convert AVI to several JPGs with ffmpeg and then, I run alpr on JPGs... this is time consuming...

How to run alpr -c eu *.jpg (doesn't work for me)?

In Linux, I tried something like this:

ls -latr *.jpg | awk '{print $9}' > file_list

alpr -c eu  `cat file_list`

but while this works for mencoder, it doesn't for alpr

Also, I could not find a method to store images with plates using aplpr -c eu something.avi 

I know this was discussed in the past... Any suggestion? | P

[Matt]

Probably the easiest way would be to modify the alpr binary (main.cpp) to read a list of files and process them individually.

For my benchmarks, I put some code together here: https://github.com/openalpr/benchmarks

The end-to-end test takes a list of images, processes them, and compares the results against a known correct answer.  The end result is percentage scores for detection, and overall accuracy.

What do you mean by storing images with plates?