When to use Fixed Learning
Advanced Fixed Learning is a precision tool for a face that repeatedly needs attention. It is not a required day-one step.
- One face is repeatedly marked Unknown or read incorrectly.
- Two values are often confused.
- A special font, finish, or orientation mark is difficult to recognize.
- The rolling-surface heatmap shows limited coverage.
Run a focused session
- Confirm the intended active Dice Set.
- Open Training → Advanced - Fixed Learning.
- Select the die type and known face value.
- Check the selected value carefully before every capture.
- Place the die naturally and wait for the die and camera image to settle.
- Move and rotate the die between samples.
- Pause whenever you adjust lighting, focus, or placement.
- Stop when the target is reached or repeat detection completes the session.
Use the heatmap as a guide
The optional heatmap beside Last Captured Dice helps distribute examples across the tray. Red areas need the most attention, yellow areas may benefit from more examples, and green areas have stronger coverage. There is no need to fill every pixel.
Avoid duplicate-heavy sessions
Small, deliberate sessions are usually more useful than many nearly identical images. Change position, angle, and rotation between captures. Reposition the die when repeat detection warns you.
Train and test afterward
Saved Fixed Learning images do not change recognition immediately. Run Training → 2. Train Recognition Model, confirm the intended model is active, and test fresh physical rolls.
Difficult dice may need several short sessions. Compare against the protected Factory Baseline whenever that helps you judge progress.
Fixed Learning troubleshooting
Capture does not begin
- Confirm the correct camera, Dice Set, die type, face value, and die count.
- Make sure every selected die shows the same preset value and is fully inside the Rolling Area.
- Move the dice, then let them become completely still.
- Review Motion threshold if the tray never settles or settles too quickly.
Repeated-image warnings appear quickly
The session is seeing nearly identical views. Move the die to another area, rotate it, change its natural resting angle, and resume. Do not defeat repeat detection by making tiny movements that add no useful variation.
The wrong face value was selected
Stop the session and use Delete Last Training Cycle when the most recent cycle was labeled with the wrong preset. Confirm the correct face before continuing. Incorrect labels can teach the personal model the wrong result.
The heatmap still contains red areas
Use the heatmap as a guide rather than a completion requirement. Add useful views in underrepresented parts of the tray, but do not fill red areas with duplicate images. Natural variety is more important than making every pixel green.
The weak face does not improve
- Run Check Training Images and remove poor or mislabeled examples.
- Complete Recognition Model Training; Fixed Learning images do not take effect immediately.
- Confirm the updated model is active.
- Test fresh natural rolls rather than fixed placements.
- Compare the result with the Factory Baseline or restore a known-good backup.
6/9 or 60/90 remain confused
Keep dots, underlines, and other orientation marks sharp and visible. Reduce glare, avoid shadows across the mark, and capture different rotations that still preserve its meaning.
