Multiple categories with various number of keypoints

[i008]

Hello, i have a case where my dataset consists of multiple classes with different keypoints (types and numer) Think of it as having a model that will do animal pose estimation for all the animal types.

How do i make it work in this framework? In my previous attempt (not using mmpose) i combined the keypoint labels so for instance
class-A: 5 keypoints
class B: 10 keypoints

And for each instance i would predict 15 keypoint maps - and just for the instances where keypoints are not relevant i would set visibility to false etc. The model would learn this association.

What would be the correct approach using mmpose and how to prepare the training_datasets and configs ?

[jin-s13]

One way might be

1. For both class-A and class-B, transform the original labels to 15 keypoints. And for the missing keypoints, just leave them as [0, 0, 0].
2. Merge the two dataset label json files into one.
   And now, you have obtained a new dataset with 15 keypoints.

[i008]

Thanks for the prompt answer and confirmation:) i will try this approach.