Object recognition and live tracking
Move from examples to located targets and ongoing observation. Connect YOLO11 + VOT tracking to your own workflow decisions.
Image matching versus object tracking
A fixed button at the same scale is different from an object changing position, angle, scale or background. JAQ Macro offers fixed-image search alongside exemplar-based tracking and trained models. Choose the method for the target’s actual appearance.
Three engine approaches
Instant tracker (VOT Alpha) starts with annotated examples and does not require a separate YOLO training run. Local search, visual features and color signatures help reacquire the target.
YOLO11 ONNX uses a prepared model with DirectML GPU or CPU inference where supported. Data, annotation and independent validation directly affect the result.
Automatic hybrid combines suitable YOLO detection with local VOT tracking and stabilized box motion. Without a model, valid exemplars can support the tracking path. Engine choice alone does not guarantee every target.
Alpha masks and GrabCut
Too much background inside the box can make the floor texture the apparent target. Mark foreground/background, inspect the transparent mask and preserve distinguishing object details. Save the corrected mask and retest.
360° Copy-Paste data
Place object examples in varied contexts. Configure rotation tolerance, shadows, background blending and color/lighting variation. Choose 0°/45°/90°/180° tolerance for realistic appearance changes.
Synthetic examples can accelerate data preparation, but do not invent genuinely unseen object details. Real scene samples and independent validation remain important.
Appearance and multiple targets
A strict appearance filter can preserve a specific look, but may reject a target whose angle changes. Evaluate unique, highest-confidence and nearest-center rules. Live tracking also has a best-target-only lock option.
Live HUD and desktop overlay
Inspect position, detected size, match score, active engine and latency/FPS. A transparent overlay can show boxes on the desktop. Values depend on the scene and hardware; an example is not a performance guarantee.
Connect observation to execution
Prepare the library resource → select Trained object in Find image → choose scope/region → preview → check Target_Found → act → verify the resulting state. The tracking box does not click by itself.
Original demo validation
The pictured world was created in code specifically for JAQ and contains no third-party game assets. Three translated positions of the sentinel were checked with the actual local detector. This is not a commercial-game benchmark or a claim that YOLO was trained for the scene.
Four questions for a useful test
- Is the correct object selected when present?
- Are there false matches when absent?
- Are similar shapes/colors distinguishable?
- What changes under motion, scale variation and occlusion?