YOLOv26 Beginner Questions: Human + Hand + Action Keypoints Annotation

Support
, , ,
1

Hi everyone,

I’m a beginner with YOLOv26 and just started exploring human and hand keypoint annotation. I have several questions regarding labeling and configuration and would appreciate any guidance from the community:

  1. Bounding box class restrictions
    Before annotating keypoints, can bounding boxes only be labeled as person and hand? Is it possible to create custom classes for action recognition?

  2. Action recognition with multiple classes
    If I want to perform simple action recognition, can I assign different classes to bounding boxes for different actions and annotate keypoints within each box?

  3. Keypoint count and YAML configuration
    Can a single bounding box contain all 59 keypoints (17 body keypoints + 21 keypoints for each hand)? If so, how should the YOLOv26 YAML file be configured for keypoints and skeleton?

  4. Human-hand association
    Do human and hand keypoints need to be strongly associated? Should left and right hands be distinguished?

  5. Simultaneous object, human, and hand annotation
    Is it possible to annotate objects, humans, and hands in the same image? If yes, what annotation rules or configuration details should I be aware of?

I’m not very familiar with YOLOv26’s annotation workflow and YAML configuration yet, so any examples, templates, or advice would be greatly appreciated.

Thanks a lot! :folded_hands:

2
  1. You can label as any class you want
  2. Yes
  3. It can contain all the keypoints if they are the same class. But if body and hand are different classes, they require separate boxes.
  4. If they are separate classes, they don’t need any association. Symmetry should be defined using flip_idx. Pose Estimation Datasets Overview - Ultralytics YOLO Docs
  5. Yes. There are no rules except you should be consistent. You can’t label hands in one image but skip labelling the hand in another image.

For annotation, you should try Ultralytics Platform

1 Like