hi,ankush-me,i want to know how to generate colors_new.cp from IIIT-5K word dataset,do you have ideas? about synthtext

[ankush-me]

The colors_new.cp is generated through k-means clustering of the word images from the IIIT-5K dataset.

[cjnolet]

I've always assumed you were clustering into 3 clusters (for foreground, background, and shadow/border) similar to how Max outlined in his dissertation. Is this correct?

…

Sent from my iPad

On May 4, 2017, at 11:59 AM, Ankush Gupta ***@***.***> wrote: The colors_new.cp is generated through k-means clustering of the word images from the IIIT-5K dataset. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[ankush-me]

I think I used k=2. From the paper:

Pixels in each cropped word images are partitioned into two sets using K-means, resulting in a colour pair, with one colour approximating the foreground (text) colour and the other the background. When rendering new text, the colour pair whose background colour matches the target image region the best (using L2-norm in the Lab colour space) is selected, and the corresponding foreground colour is used to render the text.

[doudoubean]

thank you ,I really appreciate your help. could you give a paper link or a github project link about generating through k-means clustering of the word images from the IIIT-5K dataset ? look forward to your kind reply.

[ankush-me]

@cjnolet You're right. I dug up my script, and it seems I used K=3, and picked the two clusters with the maximum "distance".
@doudoubean I have posted a gist of the code I used here.