arashhosseini/3d-pose-baseline

A simple baseline for 3d human pose estimation in tensorflow. Presented at ICCV 17.

3d-pose-baseline's Issues

What data to use from Human3.6M?

I would like to know what data to fetch from Human3.6 website. Are we supposed to download all the content of poses folder per subject in training data or a specific sub-folder ?

Error running openpose_3dpose_sandbox.py

Hello @ArashHosseini !
I have been experiencing issues to run your repository.
My operating system is Lunux Ubuntu
My TensorFlow error is 2.0, but I have tried to replace the import of TensorFlow by tensorflow.compat.v1
My python version is 3.7
The error I get when I run the openpose_3dpose_sandbox.py over the directory in which my json files are stored is the following:
File "src/openpose_3dpose_sandbox.py", line 230, in main
smoothed = read_openpose_json()
File "src/openpose_3dpose_sandbox.py", line 56, in read_openpose_json
_data = data["people"][0]["pose_keypoints_2d"] if "pose_keypoints_2d" in data["people"][0] else data["people"][0]["pose_keypoints"]
IndexError: list index out of range

This is what one of the json files contains:
{"people": [{"pose_keypoints_2d": [603, 149, 621, 188, 529, 192, 463, 297, 446, 184, 717, 184, 800, 290, 817, 176, 573, 458, 507, 622, 0.0, 0.0, 686, 454, 752, 622, 0.0, 0.0, 586, 133, 621, 133, 568, 133, 647, 133]}]}

Could you clarify the input 2d and output 3d data?

Hello, since the data is not available.
Could you clarify the 2d and 3d pose data input to the network?
I believe the 3d is root center coordinate in camera space.
Is 2d data in original image space in pixel unit?
Am I correct?

If that is the case, this model is trained on H36M specific camera configuration.
When applied on images in the wild, is that still valid considering different resolution and camera parameters?

Adding feet key points to the model for 3D conversions

Hello,
I want to use the pretrained model that you provide to convert my estimated 2D coordinates from openpose body 25 model, where the toes and heels of both feet are also included as predicted 2D coordinates, into 3D coordinates. I want to ask 2 questions:

1. Is it possible to do so, since the pretrained model was not tuned for these extra toe and heel coordinates. In fact, I would like to elaborate that I want to only use 12 coordinates in total, where 6 are feet coordinates (toes and heels) and 6 are other body coordinates that your model is trained on. So, I can append 5 body coordinates as zeros to make the input size 17 (same as your network's input and output row size) and then run your pretrained model. Would the pretrained model still be valid to use in this case.
2. If it is still feasible to use pretrained model for predicting these extra feet coordinates in 3D as well, what changes are required in the openpose_3dpose_sandbox.py or any other files as well, to give 2D input with extra 6 coordinates (2 toe and 1 heel from each foot), and get 3D output with these extra 6 coordinates estimates as well.

I want to emphasize that I do not have any ground truth labels, hence I must use the pretrained model only without any fine-tuning too. Thanks.

Can somebody provide me with Processed 3d_data.json (Human3.6M 16points)

Can somebody provide me with Processed 3d_data.json (Human3.6M 16points)?
I just want to use the data to verify my unity program……
You can give me the Download link or email me([email protected])
Thank you very much!

Unable to download Human3.6m dataset

Great work man!

I have some issues downloading the human3.6m dataset. I had sent them an email requesting for member access however the email did not exist. Would you be able to guide me where I am able to find this dataset by any chance? Alternatively, would you be able to supply the pre-trained model?

Hope to hear from you soon!

data or pretrained model links outdated

Hello,
The link for data fetch (to train model fro scratch) and also the link for the pretrained model(to evaluate) don't seem to work anymore. Would you kinldy update them ?
Thanks

Can't find data either here or in master branch

Hello.
The data file is gone and can't be downloaded.
I have checked the master and also this branch.
Could you make them available?

Thanks.

missing: 3d-pose-baseline/src/../data/h36m/metadata.xml'

Trying to get the pre-trained model to run, but this config seems to be missing.

Traceback (most recent call last):
File "src/predict_3dpose.py", line 524, in
tf.compat.v1.app.run()
File "/home/benjamin/anaconda3/envs/tf/lib/python3.7/site-packages/tensorflow/python/platform/app.py", line 40, in run
_run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)
File "/home/benjamin/anaconda3/envs/tf/lib/python3.7/site-packages/absl/app.py", line 303, in run
_run_main(main, args)
File "/home/benjamin/anaconda3/envs/tf/lib/python3.7/site-packages/absl/app.py", line 251, in _run_main
sys.exit(main(argv))
File "src/predict_3dpose.py", line 519, in main
sample()
File "src/predict_3dpose.py", line 397, in sample
rcams = cameras.load_cameras(os.path.join(this_file, "..", FLAGS.cameras_path), SUBJECT_IDS)
File "/home/benjamin/reverie/pose3d/3d-pose-baseline/src/cameras.py", line 162, in load_cameras
xmldoc = minidom.parse(bpath)
File "/home/benjamin/anaconda3/envs/tf/lib/python3.7/xml/dom/minidom.py", line 1958, in parse
return expatbuilder.parse(file)
File "/home/benjamin/anaconda3/envs/tf/lib/python3.7/xml/dom/expatbuilder.py", line 910, in parse
with open(file, 'rb') as fp:
FileNotFoundError: [Errno 2] No such file or directory: '/home/benjamin/3d-pose-baseline/src/../data/h36m/metadata.xml'

cameras.h5

Thanks for your interest in our research!

If you have problems running our code, please include

1. Your operating system
2. Your tensorflow version
3. Your python version
4. The stack trace of the error that you see

This is research code, and its primary purpose is to reproduce the scientific results that you see in our paper.
If you are looking for ways to extend our work such as

• handling arbitrary images
• handling videos as input
• using other 2d detectors
• large-scale deployment
• handling missing data
• you want to make a startup out of this code

then pull requests are welcome, but it is very unlikely we will have the time to help you.

Accuracy of 3D coordinates on own images without tuning the model

If I have computed the 2D coordinates in pixels using my own camera and then used your code to compute 3D coordinates in cm (in camera coordinates) without tuning the model, i.e. I used your code directly to infer 3D points (because in my data, actual 3D coordinates are unknown and hence, I cannot tune the model). Would this method still give me correct 3D coordinates in cm? Or will it not, because my camera is different than what was used for training the model. Thanks.

doesn't work with tensorflow 2.x

is this code still maintained or used by anyone? I would be keen to contribute if necessary.

there is no support for tensorflow 2.x. please let me know if anyone is working on that

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

might just do the trick

3d coordinate center

Thanks for your great work!

Where is the origin of the coordinate system, which contains all the 3d coordinates of all points?

About h36m.zip

Sorry to bother you!!
I want to download the h36m.zip from the DropBox link which you offer,but the link was not alive.
Can you offer me the new link to get h36m.zip? Please, I really this to study my research.
Thank you!!

about the mean and std of the 2D keypoints

I tried to predict 3D kypoints from 2D keypoints of my own images, but I can't download the dataset h36m.zip. As a result, I cannot run the the script openpose_3dpose_sandbox.py. I think I don't need the own dataset for refernece, but I need the mean and std of the dataset, so could you give the value of mean and std of the 2D keypoints?

Can3d_data.json

Thanks for your interest in our research!

If you have problems running our code, please include

1. Your operating system
2. Your tensorflow version
3. Your python version
4. The stack trace of the error that you see

This is research code, and its primary purpose is to reproduce the scientific results that you see in our paper.
If you are looking for ways to extend our work such as

• handling arbitrary images
• handling videos as input
• using other 2d detectors
• large-scale deployment
• handling missing data
• you want to make a startup out of this code

then pull requests are welcome, but it is very unlikely we will have the time to help you.

Multi-person case

@ArashHosseini
Hi, can we use that code for multi-person case?
Why i get that animation for video

404 error for h36m.zip from dropbox

Attempting to use the instructions and failing on the following step:

wget https://www.dropbox.com/s/e35qv3n6zlkouki/h36m.zip