search

MichalBusta / FASText

Efficient Unconstrained Scene Text Detector
GNU General Public License v2.0
191 stars 74 forks source link

bad segmentation in /tmp/base_chars.png #10

Closed michath closed 8 years ago

michath commented 8 years ago

hi, running tools/segmentation.py creates /tmp/base_chars.png:

thanks!!

MichalBusta commented 8 years ago

Hi,

On 08/16/2016 04:58 PM, michath wrote:

hi, running tools/segmentation.py creates /tmp/base_chars.png:

  • i was not able to determine where exactly in code it is created

src/Python/pyFastText.cpp : get_char_segmentations

  • i guess it contains the resulted segmentation, which is bad

the "chars" visualization is rubbish:

  • works just on simple images : there are all segmenations from all scales, bigger segmentations hides smaller etc.
    • i wonder if the bad result is due to an unfit cvBoostChar.xml

probably no, you can try just draw the bounding boxes of the segmentations (first 4 columns in segmentation array).

  • which API should i use to recreate cvBoostChar.xml myself?

no API provided.

All the best, Michal

  • any code example for such call? thanks!!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/MichalBusta/FASText/issues/10, or mute the thread https://github.com/notifications/unsubscribe-auth/AD6jsA6wEZLg18jvGtetkc0t-PHmAz5Sks5qgc_5gaJpZM4JlgNv.

michath commented 8 years ago

thanks, but - array (char) segmentations has ~2000 rows per source image given in figure 11 (top row) of the paper. how can this number be narrowed down to produce the resulting text in figure 11 (bottom row)? [i guess this large number explains /tmp/base_chars.png] thanks again :)

MichalBusta commented 8 years ago

On 08/17/2016 10:26 AM, michath wrote:

thanks, but - array (char) segmentations has ~2000 rows per source image given in figure 11 (top row) of the paper.

correct, see Table 2 for expected numbers. the FASText key-points is very low preprocessing stage, it do the job well: sample the characters segmentation -> low cost approximation of CSER

how can this number be narrowed down to produce the resulting text in figure 11 (bottom row)? [i guess this large number explains /tmp/base_chars.png]

text clustering and classification.

thanks again :)

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/MichalBusta/FASText/issues/10#issuecomment-240346582, or mute the thread https://github.com/notifications/unsubscribe-auth/AD6jsDukOyrhpdCl5vgbwCZ0wLeCf6h0ks5qgsWxgaJpZM4JlgNv.

michath commented 8 years ago

thanks, but:

thanks a lot!

MichalBusta commented 8 years ago

On 08/17/2016 04:50 PM, michath wrote:

thanks, but:

  • where in code do i get the features for such classifier?

CharClassifier.cpp -> extractCharFeatures

  • the paper mentions four features, but cvBoostChar.xml has six.

yes, the code evolved since publication deadline and release time. If desired, I can publish change list. (there are also small changes in pattern cheking, etc ... )

  • where in code do i get the output of AdaBoost? is it applied? thanks a lot!

the output of classifier is in segmentations array: see quality -> returns the np array where row is: [bbox.x, bbox.y, bbox.width, bbox.height, keyPoint.pt.x, keyPoint.pt.y, octave, ?, duplicate, quality, [keypointsIds]]

there is no filtering of the segmentations, you can just query for quality: segmentations[:, 9] > q

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/MichalBusta/FASText/issues/10#issuecomment-240436685, or mute the thread https://github.com/notifications/unsubscribe-auth/AD6jsABy6kWBGTQI-Lsbyio08GqPNNtEks5qgx_SgaJpZM4JlgNv.