argman / EAST

A tensorflow implementation of EAST text detector
GNU General Public License v3.0
3.02k stars 1.05k forks source link

Slow executation #242

Open anavc94 opened 5 years ago

anavc94 commented 5 years ago

Hi,

I am getting extremely good results using EAST on different images. First of all, congratulations! The negative part I am finding is that EAST is slow, even running on the GPU GTX 1060. For me, it can take like 400 or 500ms for computing the text detection in one SD image using a GPU , so it cannot run in real time. Is there a way to improve performance on speed? Thanks

DhavalNiphade commented 5 years ago

Have you figured out why?

I'm running it on an Intel Xeon @3.2Ghz with Nvidia RTX 2080Ti and it's taking 2-3 seconds per frame. Tried it with a compiled version of TensorFlow for my system configuration too.

arvindrajan92 commented 4 years ago

Were you able to resolve this issue? @DhavalNiphade @anavc94 @argman @zxytim

anavc94 commented 4 years ago

Hey @arvindrajan92 ,

in my case the slower part was de nms algorithm, maybe 'cause I couldn't compile lamns on windows. I just changed that function.

Regards, Ana

arvindrajan92 commented 4 years ago

Hi @anavc94,

Thank you for sharing your findings. I agree with you, it is the nms part of the algorithm that is slowing things down. The more the texts in the image, the longer it takes. In my case, I replaced the nms algorithm with a faster one as described (with source code) here. This speeds up the detection quite significantly.

Cheers, Arvind

SabraHashemi commented 3 years ago

@arvindrajan92 @anavc94 @arvindrajan92

what was your final execution time for single line text image ?