google-deepmind / temporal-3d-pose-kinetics Goto Github PK

Your code is wonderful!

I have a question. The google driver link of Kinetics (325 G) provided in this repo is referred to original Kinetics-400? If so, I want to download this dataset for action recognition experiments.

Thank you in advance.

Hi,
I loved the work you have done and its quite interesting.
I had two doubts -
1)How to run it for our own dataset containing videos.
2)How to extract the final 3D-joint locations(after the bundle adjustments)?

Hi there,

I am trying to replicate this paper's work which is why I am using the custom regressor that is provided. However, I am unable to generate the correct results.

In the image above, the yellow points are the ground truths, the blue points are generated using the Cocoplus Regressor provided in the HMR GitHub repository and the red points are generated when I use the Custom Regressor that you provided. I just used the Custom Regressor in place of the Cocoplus Regressor without changing any code (as it is a drop-in replacement).

I am not quite sure about the issue behind getting the wrong results. Is there anything that needs to be altered that I am missing (for eg: something is normalized that needs to be changed before usage)? I will be looking forward to hearing back from you. Thanks!

Note: I have tried using other joint regressors as well, and those work fine.

Thank you for your great works!

Could you release your pre-trained models like HMR did?

The descriptions from HMR Github:

...
Download the pre-trained models

wget https://people.eecs.berkeley.edu/~kanazawa/cachedir/hmr/models.tar.gz && tar -xf models.tar.gz

...
python -m demo --img_path data/coco1.png
...

And then,
we can generate results from our input videos by your paper.

Hi @HMPH and @cdoersch,

Thanks for releasing the dataset. Is it also possible to release the 2D keypoints used for fitting, too? Thanks!