Object library
Collect examples, annotate the object, choose a suitable engine and test tracking in the actual scene.
Fixed image or object?
For a stable interface button, Find image → Fixed image is often the most direct start. Consider object-library examples and training when appearance, angle or background changes.
1 · Collect examples
Give the object a clear name and group. Select the target window or screen scope, capture different views and annotate the object tightly. Separate the target’s appearance from unnecessary background.
Include varied backgrounds and scenes without the object as appropriate. Incorrect annotations are not corrected simply by training longer.
2 · Prepare and train
Choose a data strategy for the selected object or scene. Rotation and Copy-Paste augmentation should reflect real variation. Select input size and training settings for the hardware and target scale.
With an appropriate YOLO11 ONNX model, the hybrid engine combines detection and local tracking. Without a model, exemplar-based local tracking can be available. These methods do not guarantee equal results in every scene.
After adding samples, check the training state. Do not assume an existing model automatically includes the new data. Training metrics do not replace testing in the target application.
3 · Live tracking
Select the target and object, then observe first. Is the right object selected? Are similar objects confused? Does the result disappear when the target is absent? Test changes in angle, scale, lighting and occlusion.
Connect it to a workflow
In Find image, choose Trained object and the library resource. Define the multi-detection rule, test the search, then add a click or conditional continuation.
Backup scope
PNG workflow templates and the local training/ONNX library are different resource scopes. Do not assume every local model is included in a .jaqb backup; check the required resources before moving to another machine.
Continue with the tracking guide
Explore Object tracking for the VOT, YOLO11 and hybrid differences, Alpha masks, synthetic data and live HUD. The original demo scene uses real local detector results and is not a commercial-game benchmark.