focused-window-detector (RF-DETR nano, CoreML)

Single-pass RF-DETR (nano) that detects macOS windows and the cursor from a live screen capture. Exported to CoreML for on-device inference on Apple Silicon.

Classes: focused_window, unfocused_window, cursor.

Strong at separating focused vs unfocused windows in real time.

Files

rf-detr-nano-checkpoint_best_total-2-fp32.mlpackage/   # CoreML model

Input / output

• Input: image 384 x 384 (RGB, scale 1/255).
• Outputs: var_2267 = boxes [1, 300, 4] (cxcywh, normalised), var_2270 = logits [1, 300, 4].

Usage

Runner (Swift + ScreenCaptureKit live overlay): https://github.com/cianmcnally/focused_window_detector

hf download Cianmcnally/focused-window-detector --local-dir ./model

Known limitations

• Cursor recognition covers only a narrow range of cursor shapes/sizes.
• Cursor bounding boxes are loose (not pixel-tight).
• The Dock is sometimes misdetected as unfocused_window on hover.

These are training-data gaps; see the runner repo for the improvement plan.