Fasten in cpu

Question

Fasten in cpu

Zumbalamambo opened this issue 7 years ago · comments

It is laggy in cpu. how can I optimize it?

Tim Ho · Answer 1 · Sat Aug 05 2017 14:42:19 GMT+0800 (China Standard Time)

You might need a Nvidia GPU that supports CUDA to run in realtime. For CPU, check if Tensorflow provides any optimization for it. And of course you can reduce the input image size or train a smaller model by yourself to get a speedup.

Also in the code, find the line "sess.run([model.stage_heatmap[5]]...", you can adjust the number from 0 to 5 which means how many stages are used. This is a trade-off between speed and accuracy.

Zumbalamambo · Answer 2 · Sun Aug 06 2017 15:23:07 GMT+0800 (China Standard Time)

Can you please tell me how I can train my own model?

justinluyao · Answer 3 · Fri Aug 11 2017 14:45:23 GMT+0800 (China Standard Time)

hi !

thanks for your work！

can you tell me what kind of data you used for training?

Tim Ho · Answer 4 · Fri Aug 11 2017 19:48:08 GMT+0800 (China Standard Time)

@Zumbalamambo
You need to feed placeholder with "input image", "center map", "joint heat map", please read the paper for more detail. I will release my training script after some refactors.

@justinluyao
I manually labeled thousands of images with 2D joint locations, and generate some 3D model images from LibHand. If you find publicly available dataset please let me know too.

justinluyao · Answer 5 · Fri Aug 11 2017 23:10:53 GMT+0800 (China Standard Time)

do you mean you generate the hand images with LibHand and label the joints positions on the generated images by yourself?

Tim Ho · Answer 6 · Sat Aug 12 2017 00:23:21 GMT+0800 (China Standard Time)

LibHand can not only generate images but also with joints labels, by projecting 3D model joints onto 2D image spaces.

sarahzhouUestc · Answer 7 · Mon May 21 2018 10:56:33 GMT+0800 (China Standard Time)

@timctho Hi timctho, thanking for you sharing of your project. I am doing my training with it. However I have some confusion about it, could you please do me a favor?

Need I feed background channel? cause I noticed that shihenw uses 4th channel which represents the background, and its values are all 0. actually I don't know why need to do this. While in your project, the input image still have 3 channels, right? and I just need to generate background heatmap, then the total number of groundtruth heatmap is (nun_of_joints + 1), am i right?
The center of center_map is the center of torsobox, right? but what is the sigma of it , how can I determine it?

Thank you very much!

Tim Ho · Answer 8 · Mon May 21 2018 23:17:32 GMT+0800 (China Standard Time)

1.For original CPM paper I think the input image only have rgb channels. And yes you need to generate the background heatmap which is complement to all the joint heatmaps. 2.Yes. In hand project, I remember the sigma is 7 or nearby. You can play with it a little, while I find out little difference. Best wishes, *Tim Ho* (何哲廷)

…

On Mon, May 21, 2018 at 10:56 AM, sarahzhouUestc ***@***.***> wrote: @timctho <https://github.com/timctho> Hi timctho, thanking for you sharing of your project. I am doing my training with it. However I have some confusion about it, could you please do me a favor? 1. Need I feed background channel? cause I noticed that shihenw uses 4th channel which represents the background, and its values are all 0. actually I don't know why need to do this. While in your project, the input image still have 3 channels, right? and I just need to generate background heatmap, then the total number of groundtruth heatmap is (nun_of_joints + 1), am i right? 2. The center of center_map is the center of torsobox, right? but what is the sigma of it , how can I determine it? Thank you very much! — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AFYMF0tOg8HVhoVIMUaEbbW_z_XJvd_rks5t0izigaJpZM4OuWRo> .