What is keypoint annotation?
Keypoint annotation consists of various particularly numbered points called keypoints which are connected by edges. They are a great way of tracking variations between objects which always have the same structure (e.g. human figures and facial features). The output are the x, y coordinates of the numbered keypoints always in the same order.
Where can it be used?
Keypoint annotations are very well suited for movement tracking and prediction, human body parts detection, gesture and facial recognition, pose identification for AR/VR, or even in sign language transcription.
- Medical: tracking instruments in robotic-assisted surgery
- Geospatial: monitoring crane movement on construction sites
- Automotive: tracking the movement of pedestrians on the street
- Industrial: detecting hand gestures of manufacturing workers
- Agriculture: tracking the movement of livestock
- Retail: detecting shopper actions in smart supermarkets
- Once the skeleton is set up on the annotation tool, the only thing needed is to adjust the nodes and annotation becomes much easier.
- Keypoints are helpful for annotators not to forget any node from the sequence.
- Objects must have the same regular structure in order for this annotation type to be useful.
- Edge cases where some keypoints are not visible (e.g. human is turned to the side) may produce confusion in the model.
- When defining the annotation instructions, determine whether hidden keypoints need to be deleted or their position has to be deduced by the labeler.
- You can define certain proportions (e.g. expected edge length) so as to make sure your model does not detect keypoints that fall outside of what is common sense.
- It is important to set clear guidelines on which side should be considered “left” and “right” (e.g. should annotators label the data from their point of view or from the object’s) because errors in interpretation might frequently happen.