Hello Thank you for the work. Is there a fast way of getting estimated animation as .fbx (or any other format)? To import to 3d software.

Hello,

Can you guide on how to process in the wild images ?

Thanks

Hello,

in the documentation you say that when prediction 3d coordinates, the X and Y are "aligned to input image".
Could you please explain this in detail? Does it mean that the model outputs something in the range [-1, 1] or [0, 1] and you rescale it based on the image size?

If I wanted to rescale them in the range [-1, 1], should I just divide them by the image width and height respectively? Or do you perform a square cropping (e.g., 256x256) and one should divide the output accordingly?

Also, does [0, 0] correspond to the top left corner of the image?

i use the pose&shape label get the vertices ,then i want to use the parm_cam to render the mesh in this camera view,like SPIN redner,but i make the [s,x,y] to [x,y,2×5000/224×s],but the render view is not true

can you help me?

hello,
I really appreciate this work, thanks very much.

And I have a question here, some data provide by eft is come from posetrack, and there are two version of posetrack, that is posetrack2018(82G) and posetrack2017(34G). I want to know which version of posetrack is used here, it's a hard work to download all of the data.

First, I encountered the bug 'ImportError: cannot import name 'ModelOutput''. Then, I modified the code from 'from smplx.body_models import ModelOutput' to 'from smplx.body_models import SMPLOutput' in smpl.py. Finally, I encountered the bug
OpenGL.error.GLError: GLError(
err = 1282,
description = b'invalid operation',
baseOperation = glGetUniformLocation,
cArguments = (3, b'ModelMat\x00'),
result = -1
)
and I don't know how to solve it. Could you please help me?

sorry to bother,i cant find the fine-tuning part in the demo_bodymocap.py.(T T)

Hi, thanks for the great work.

Could you please hint me how to convert from rotation matrix representation (Bx3x3) to axis-angle representation (Bx3) (used in original SMPL)?

Thank you!

What‘s “gt_keypoint_2d” means in EFT Fitting data？The first 25 are openpose keypoints, and the last 24 are smpl keypoints?

hey!

I wonder what are the approx fps for webcam atm? would be useful to know for gui mode on and off especially

Thanks for releasing the fitting code! I found that the implementation of weakProjection_gpu is "not" correct.
In simple terms, in your implementation, the projection is computed as s*pts + t. In the original HMR/SPIN implementation, it implies that the projection should be s*(pts + t). In this way, this projection can be perfectly converted to perspective projection by setting the camera translation to [tx, ty, 2*f/s * img_res].

https://github.com/MandyMo/pytorch_HMR/blob/7bf18d619aeafd97e9df7364e354cd7e9480966f/src/util.py#L117

Anyway, it a fitting procedure, the network's parameters will be updated to output the correct t under this projection.

Hello,

In the demo there is dual person prediction despite there being only one person. I'd appreciative any pointers as to how to address that issue.

Thank you.

Hello, Dr Joo.
First of all, thanks for the impressive and inspiring work that your team have made :-)
Since one of the main applications of EFT is a 'post-processing', would it be possible for you to release the pre-trained model weight of SPIN, which is trained on your own loss formula?
Definitely I could get the pre-trained SPIN model weight via SPIN GitHub, but as your paper suggest the better formula to calculate the 2D joint reproduction loss, I'd like to request the pre-trained SPIN weight on your way if it doesn't bother you that much!

Thanks for your great work!
As we all know ,the mesh result of PE may not fit the origin videos well,so i want to use EFT as Post-processing method to get better result.
So if i already have the 2d keypoints and smpl parameters got from PE,how can i use EFT to Fine tune?

Hi, I want to render SMPL mesh with perspective projection. How to set a perspective camera?

Hi thank you for sharing this great dataset first!
But I just found that annotId doesn't match the image's name in COCO annotation provided (both all and part)
I think they should match each other? The smpl param I parsed seems do not match well either.
Or did I miss something?

The annotation from COCO:

The annotation from eft:

The smpl param parsed by me:

Hi, are you willing to release the fitting code and the fine-tuned SPIN model？

Can you release the training code on the EFT dataset? I would be grateful if you could provide the code.
Have a nice day :)

Hi, thanks for releasing the training code! Recently, I want to modify this code to support smplx-based fitting by introducing extra hand or facial landmarks, could u share the JOINT_REGRESSOR_TRAIN_EXTRA_SMPLX for smplx-based fitting? It seems that you have already tried similar things.

when I run the demo visEFTFit.py , my parser definition is:

parser = argparse.ArgumentParser()
parser.add_argument('--img_dir',default="../demo/images/demo.png", type=str , help='Folder path where input image files exist')
parser.add_argument('--fit_data',default="../eft_fit/MPII_ver01.json", type=str, help='EFT data json fortmat')
parser.add_argument('--smpl_dir',default="../extradata/smpl", type=str , help='Folder path where smpl pkl files exist')
parser.add_argument('--onbbox',action="store_true", help="Show the 3D pose on bbox space")
parser.add_argument('--rendermode',default="geo", help="Choose among geo, normal, densepose")
parser.add_argument('--render_dir',default="render_eft", help="Folder to save rendered images")
parser.add_argument('--waitforkeys',action="store_true", help="If true, it will pasue after each visualizing each sample, waiting for any key pressed")
parser.add_argument('--turntable',action="store_true", help="If true, show turn table views")
parser.add_argument('--multi',action="store_true", help='If True, show all available fitting people per image. Default, visualize a single person at each time')
args = parser.parse_args()

I get error like this:

freeglut (foo): fgInitGL2: fghGenBuffers is NULL
Traceback (most recent call last):
File "E:/human_pose_estimation/eft/demo/visEFTFit.py", line 436, in
renderer = getRenderer(args.rendermode)
File "E:/human_pose_estimation/eft/demo/visEFTFit.py", line 52, in getRenderer
renderer = meshRenderer.meshRenderer()
File "E:\human_pose_estimation\eft\renderer\meshRenderer.py", line 37, in init
glRenderer.init(self, width, height, name, 'geo', color_size, ms_rate)
File "E:\human_pose_estimation\eft\renderer\glRenderer.py", line 34, in init
glEnable(GL_DEPTH_CLAMP)
File "D:\ProgramFiles\Anaconda3\lib\site-packages\OpenGL\platform\baseplatform.py", line 415, in call
return self( *args, **named )
File "src\errorchecker.pyx", line 58, in OpenGL_accelerate.errorchecker._ErrorChecker.glCheckError
OpenGL.error.GLError: GLError(
err = 1280,
description = b'\xce\xde\xd0\xa7\xc1\xd0\xbe\xd9',
baseOperation = glEnable,
cArguments = (GL_DEPTH_CLAMP,)
)

Hi,
Thanks for a great job!
Is there any plan to release to code for the fitting procedure? (or I may have missed it?)
Best regards,
-Thibault

I have downloaded the json formats data. According to my test, the length of COCO2014-Part-ver01.json is 28,062 and that of COCO2014-All-ver01.json is 74,834. This is not consistent with the picture .Maybe there's something wrong about the name of the file.

Hi, I notice that the repository only renders results with OpenGL, could you give me some suggestions on how to use pytorch3d to visualize the output results?

Hi @jhugestar,

Thanks for releasing the EFT code and data! It is much appreciated.

It seems bodymocap.datasets and from bodymocap.smplify are missing. Could you provide those packages as well?

Thank you for your release!

Do you get smpl parameters for 3DHP using Mosh? Could you share these parameters please?

Thank you again!

Hello,

While trying to run the demo there it looks for the modelViewer which isn't given. Please look into this. Thanks