How to finetune Voxelpose?

[Camila641]

trafficstars

Thanks for the great work! I'd like to finetune Voxelpose to get the 3d pose estimation from multiple views but I'm new to this. May I ask if there is any guide I can follow? I have read the repo but not sure which to follow.

Many thanks!

[liqikai9]

@Camila641 Hi. Thanks for using mmpose.

You can refer to this config , which train VoxelPose on Panoptic dataset. You can modify the corresponding settings according to your own dataset.

[Camila641]

@Camila641 Hi. Thanks for using mmpose.

You can refer to this config , which train VoxelPose on Panoptic dataset. You can modify the corresponding settings according to your own dataset.

Thanks for the quick response! To finetune it on my custom dataset, should the dataset in Panoptic format and are 3D keypoint coordinates needed? I'm a little bit confused about how to get the 3D keypoint coordinates.

[liqikai9]

should the dataset in Panoptic format

Yes. The images and camera calibration data should be formatted into the Panoptic style.

are 3D keypoint coordinates needed?

Yes, the ground truth 3D keypoint coordinates annotations are necessary.

[Camila641]

Other than #1679, Do you have any recommended ways or tools to get the 3D keypoint coordinates annotations?

[liqikai9]

I think you may try labelbee.

Or you may read the detail about the Panoptic dataset, and see if there are some guidances to get 3D keypoint coordinates annotations.