What a label means
A label is the name of something the camera should recognize, such as Robot, Hand, or Background. Choose names that describe what is actually visible in the frame.
Choose distinct labels
Start with labels that look clearly different from one another. If two labels use nearly identical scenes, the prediction will be harder to interpret.
Capture varied examples
Show the object at slightly different angles, distances, and positions. Keep those variations realistic: the training set should resemble how the camera will be used in the finished project.
Add a background label
Capture the scene without the target object. This gives the model a clear example of the state in which no action should happen.
Test before adding hardware
Start prediction and move the object through the frame. Check that the label changes at the right time and that confidence remains reasonably stable.
Improve a weak label
- Add new angles instead of repeating the same frame.
- Remove clutter that appears in only one label.
- Improve the lighting.
- Keep the camera position consistent with the final project.
- Recapture examples after a major scene change.
Tip