search

chaneyddtt / Generating-Multiple-Hypotheses-for-3D-Human-Pose-Estimation-with-Mixture-Density-Network

Code for our CVPR2019 paper: Generating Multiple Hypotheses for 3D Human Pose Estimation with Mixture Density Network
MIT License
97 stars 9 forks source link

About the missing "camera.h5" file #17

Open KVBK01 opened 4 years ago

KVBK01 commented 4 years ago

Thanks for your great work!

After a careful look, I know that the pre-processing of Human3.6M part is done by una-dinosauria, but unfortunately, in the current version of the una-dinosauria's code, the "camera.h5" file is removed (and now is substituted with another file), thus it is no way to successfully run your code since it requires this file...

I am wondering would it be possible that you can share me with the "camera.h5" so that i can run your code?

Best regards!

chaneyddtt commented 3 years ago

Hi @Vsilent01 I can not share with you the file because of the license issue. Actually you can still follow una-dinosauria's current preprocess code to generate the 2d ground truth joint locations. They use the metadata.xml instead of the camera.h5 to obtain the camera parameters in their current code, and you can download the metadata.xml from the official website. let me know if you have further questions.

KVBK01 commented 3 years ago

Hi @Vsilent01 I can not share with you the file because of the license issue. Actually you can still follow una-dinosauria's current preprocess code to generate the 2d ground truth joint locations. They use the metadata.xml instead of the camera.h5 to obtain the camera parameters in their current code, and you can download the metadata.xml from the official website. let me know if you have further questions.

Thanks for you reply! May i ask one more question about the "stacked-hourglass" part? Now the code released by una-dinosauria only supports training from groundtruth. Can you share me with the trained stacked-hourglass file so i can have train the network from stacked-hourglass detections? Or maybe other ways to obtain the trained stacked-hourglass results?

Thanks again for your patience !

chaneyddtt commented 3 years ago

What I would suggest is that you can take the pretrained model from @crockwell and generate the 2D detections. Note that you will have to finetune the pretrained model on the human3.6m dataset to get more accurate 2D detections. You can also directly use the 2D detections from the pretrained model, and you will not get as good results as we reported in our paper in this case.

KVBK01 commented 3 years ago

What I would suggest is that you can take the pretrained model from @crockwell and generate the 2D detections. Note that you will have to finetune the pretrained model on the human3.6m dataset to get more accurate 2D detections. You can also directly use the 2D detections from the pretrained model, and you will not get as good results as we reported in our paper in this case.

Thanks again for your kind reply! My question is well resolved!