mandymo / pytorch_hmr Goto Github PK

The data link you provided seems to have expired. Could you please share it again? Thanks a lot.

Hi,

would you like to offer the process scripts of datasets?

downloading the datasets from Baidu Netdisk is much too slow

Thanks for your reproduce! I also try to reproduce another work --- estimate human shape from only one color figure. Some part is similar to HMR. If you can share your test code, that will be a great apreciate!

Is there a way to apply 3D Mesh Inference on given image?

Great work!
Can you please update the license file like MIT or BSD.

Thanks.

I notice that the share of model.zip has been canceled. Can you re-upload the file? Thx!

You mentioned that in order to train the model, model.zip should be unzipped. However, I cannot find model.zip in your repo. Could you please upload it as soon as possible? Thanks.

Hi, it's me again.

So far I can train the model using your scripts, but still have some problem processing the outputs.

I wrote a script almost the same as the original demo.py, replacing the tensorflow inference code with forward functions in model.py, and processed the 2d kps (output range is (-1,1)) to be drawed.

But I can't get the right output rendering image.

In fact, even when I run with the same weights file twice, I can't get the same outputs.

I'm new to pytorch, maybe I have encountered some Novice problem。

Ps: I use the copy_state_dict funtion to load weights file.

Which Python version did you use?

I'm getting an error for trainer.py line 302

  for iter_index in range(config.args.iter_count):
            try:
                data_2d = loader_2d.next() # <-

next is a built-in function from python 3 and should be called next(loader_2d)

And with Python 2.7 I get an error for copying a list in COCO2017_dataloader.py at line 117

self.boxs.append(box.copy())

This feature is only from Python 3.5

Since human3.6m is not available, how to use mosh data provided by the original tensorflow version for adv training?

I have been able to get everything working except the discriminator.

How do you train the discriminator? As far as I can tell, the code only loads the trained model.

Hi @MandyMo,

Could you please upload all the files in google drive / dropbox?

In order to create a Baidu account - to download the files, a chinese mobile number is needed which I don't have.

the shared data has been can cancelled. I have already down load the videos of h3.6m, but I don't have the annot.h5 file. Anyone here can share this file to me? thanks a lot.

Hi Xiong,

Thanks for sharing this PyTorch port!
I was wondering, can you share the pretrained PyTorch weights with me?

Hi, have you totally reproduced results in the paper (comparable number for evaluation metrics)? Thanks.

I try to train the model ,but I can't get the h3.6m dataset. I try to only use other datasets,but the result is weird. I tried to contact the admin of h3.6m , but I was refused . Can you share the dataset with me via BaiduNetdisk or email [email protected]? By the way ,I am a grade student of UESTC. Thank you so much !!!!!!!!

Hi,

I cannot seem to download the baidu files. I even tried the chrome extension. Is there a workaround?

Hello @MandyMo ,
Thank you for releasing the code, it is a really great work.
I trained with this code and got reasonable results on 3d datasets, but terrible results on 2d datasets. The losses all seem stable. Then I tried to change hyperparameters, but the results on 2d datasets were far from good. I don't know whether it's normal or not. Really confused.
How is the performance of your pre-trained model?
What's your performance on training set and testing set of 2d datasets?
Can you share the hyperparameter setting or give some instructions?
Thanks a lot! Following are my training details and results.

This is the default parameters setting from this code, just changed the batch_size_(2d, 3d, adv) to fulfill my GPUs. I trained 7M samples for each dataset, you can regard it as about 1M iterations with batch size of 7.
Losses:

Training results from 3d datasets, they seem reasonable.

Training results from 2d datasets, they seem terrible, but the loss for 2d points keeps stable.

Then I amplified the ratio of 2d loss for 10 times, trying to lower 2d loss. But the result didn't goes the way I wanted. The 2d loss went a little bit lower if it divides 10 comparing the former one, it still much higher than 3d keypoint loss. And the results on 2d datasets were not satisfying.
The losses:

 Training results from 3d datasets:

Training results from 2d datasets:

The BaiduNetdisk link of h36m dataset is not work yet. Can you provide the dataset again? Thank you very much!

Hi, I am interested in the 3D reconstruction of human body. Can you share the human3.6m moshed data, like shapes and poses? Thank you very much!

I can't download the file, mosh_annot.h5. Could you please tell me how to get this file? Or share it with me?

Hi, in the paper, I find the the parameters contains global rotation R (3x3), however I can't find it in the code. Do I miss something?

I can't download human36m dataset by the link you provided, and the official website is too damn slow. So can you provide the link again?

HI,
I want to run a demo.py to test trained model. However, the result turns out like this,
"finished create the encoder modules...
parameter smpl.v_template not found
parameter smpl.shapedirs not found
parameter smpl.J_regressor not found
parameter smpl.posedirs not found
parameter smpl.joint_regressor not found
parameter smpl.weight not found
parameter smpl.e3 not found"
Did I miss any code as I only use"smpl = model.smpl" ?

AI challenger keypoint dataset is not available now, can you share another copy?

Thanks for your sharing. I have downloaded official SMPL model.
But I can't find any model with coco joint plus.
Could you please tell me?

Hello, I want to run this code. How about the performence of this implementation?

thank you for your works, i would like to run the code in real time using my webcam to test the result.
how can i do that?
thanks

Hi, could you please share some extra data like camera parameters (focal length, principal pt) and translation vector so that I can render the human3.6m 3D model to the corresponding images. Thanks.

你好，我用hmr网络训练hand reconstruct任务，freihand数据集，回归cam参数+MANO参数，但是cam参数并不收敛，尤其是scale，基本不下降。
是训练需要哪些trick吗？
还是要预训练，还是分两个网络分别训练吗？

The model you uploaded to BaiduYun has been deleted. So if possible, could you share is again? We would appreciate it. Thanks!

Hello,

I'd like to use HMR for a personal project. Would you happen to have a trained model or do I have to put my GPU to work ? ;)

Thanks !

Human 3.6M didn't provide SMPL parameter. Could you explain how to calculate groundtruth parameter?

Dear @MandyMo, Thank you for your great work on the HMR pytorch implementation. I wonder if it is possible to release your pretrained model to ease my reproduction of results in the original paper. Thank you very much.