Incorrect bounding box for characters

[fireton]

Environment

• Tesseract Version:

D:\Tesseract32>tesseract.exe --version     
tesseract v5.0.0-alpha.20210506                                                                                                                                                                                     
leptonica-1.78.0                                                                                                                                                                                                    
libgif 5.1.4 : libjpeg 8d (libjpeg-turbo 1.5.3) : libpng 1.6.34 : libtiff 4.0.9 : zlib 1.2.11 : libwebp 0.6.1 : libopenjp2 2.3.0                                                                                  
Found AVX2                                                                                                                                                                                                         
Found AVX                                                                                                                                                                                                          
Found FMA                                                                                                                                                                                                          
Found SSE4.1                                                                                                                                                                                                       
Found libarchive 3.5.0 zlib/1.2.11 liblzma/5.2.3 bz2lib/1.0.6 liblz4/1.7.5 libzstd/1.4.5                                                                                                                           
Found libcurl/7.77.0-DEV Schannel zlib/1.2.11 zstd/1.4.5 libidn2/2.0.4 nghttp2/1.31.0 

• Commit Number: I'm using Mannheim build
• Platform: Windows 10 64 bit, Tesseract compiled for 32 bit

Current Behavior:

Language: Russian + English.
I'm using CAPI and trying to get separate characters bounding boxes through iterator (TessPageIteratorBoundingBox(RIL_SYMBOL,...). For attached image I get correct characters for whole line but bounding boxes are wrong.

piece_oreshek_small.zip

Line is «орешек).»

Up until second «е» character everything is great. I receive correct bounding boxes for «о», «р», «е» and «ш».

But this I get for second «е»:

This one for «к»:

This for «)»:

And this for «.»:

Obviously bounding boxes are wrong.

Expected Behavior:

Proper bounding boxes.

Suggested Fix:

I'm not qualified enough.

[zdenop]

First of all - how is your report different for other reports in issue tracker e.g. #3105? Did you try suggest solution?
Next:

• show us code that demonstrate you problem
• give info about langdata you used

[fireton]

I've tried to initialize Tesseract in legacy mode as suggested, but it gives me error. TessBaseAPIInit2 returns -1 and if I try to run tesseract.exe with --oem 0 I get this :

Error: Tesseract (legacy) engine requested, but components are not present in D:\Tesseract32/tessdata/rus.traineddata!!
Failed loading language 'rus'

I'm using Delphi wrapper all CAPI functions are incapsulated into some Delphi classes. So I can give you code but it will be not C API code. Basically I'm using TessBaseAPIInit2 to initialize engine, then pixReadTiff and TessBaseAPISetImage2 to set image, then TessBaseAPIRecognize, then I get iterator by TessBaseAPIGetIterator and TessResultIteratorGetPageIterator (btw couldn't find any info why they are different since they share same handle as it in tess code). Then I use TessPageIteratorNext(..., RIL_SYMBOL) and TessPageIteratorBoundingBox(..., RIL_SYMBOL,...) to iterate results symbol by symbol and get their bounding boxes. Also in this cycle I use TessPageIteratorIsAtBeginningOf and TessPageIteratorIsAtFinalElement to determine current position inside text line and paragraph.

What kind of info of langdata do you need?

[stweil]

The Windows installer installs "fast" models, but for legacy mode you need different ones. So please replace your local rus.traineddata by the one from https://github.com/tesseract-ocr/tessdata/raw/master/rus.traineddata (and maybe similar for eng.traineddata and other models which you use).

[fireton]

Thanks, I'll try it.

Also I've tried to run tesseract.exe with makebox config and on this same tiff it gives me correct bounding boxes. I don't get it. :(

[fireton]

Tried langdata you suggested and legacy mode. It works in terms of proper bounding boxes but recognition quality of text is poor.

Many cyrillic characters are wrong, text division (by line and blocks) also not good.

[wollmers]

Tried langdata you suggested and legacy mode. It works in terms of proper bounding boxes but recognition quality of text is poor.

Recognition quality depends on the training of the models. With --oem 0 you get better character bounding boxes, with the neural net models you get better recognition. If you need both, you should run both (As I do.)

Many cyrillic characters are wrong, text division (by line and blocks) also not good.

Then the model is maybe not well trained on the used font.

[fireton]

If you need both, you should run both (As I do.)

Tried combined mode. Same result as default. I mean wrong bounding boxes and good recognition.

[fireton]

I've read code for .box file renderer and found that it uses LTRResultIterator which is not exposed in CAPI.

[amitdo]

If you need both, you should run both (As I do.)

Do you mean oem 2 (or 3), or running Tesseract twice and somehow combine the results yourself?

[wollmers]

If you need both, you should run both (As I do.)

Do you mean oem 2 (or 3), or running Tesseract twice and somehow combine the results yourself?

Twice = 2 passes.

See #3105 (comment).

When I need correct character bounding boxes, the I run 1 pass with --oem 0, and for good recognition 1 pass with --oem 3. In my case I need the bboxes for font identification and for not or wrong recognised characters or symbols.

[woodjohndavid]

See #3477 for some further information on bounding box computation in legacy and LSTM modes.