GREENPOINT / REAL AGRICULTURAL INSTANCE DETECTION

Measure the recovery.
Keep the failures visible.

Forty image-disjoint public PhenoBench confirmation images. Parameters selected on a separate twenty-image calibration set. No GreenPoint model training.

60.2% → 60.7%weed recall · previous → selected
96.7%selected weed precision
PASS90% precision / 60% recall / crop errors ≤5%

Research prototype; not a deployed feature or independent-farm validation. Author checkpoints may have used official validation images for model selection. All predictions require human review.

Selected source with real model boxes

Crop: sage · Weed: coral · Uncertain: amber. Object matches require one-to-one IoU ≥0.5. Examples are selected after evaluation; use the complete table below for the result.

The same source-pixel crop compared across annotations and models

Complete paired confirmation

MethodCorrect / all weedsRecallPrecisionCrop → weed
previous_whole89/19645.4%87.3%0/343
previous_hybrid118/19660.2%76.1%0/343
official_yolo153/19678.1%68.3%7/343
calibrated_yolo119/19660.7%96.7%0/343
selected119/19660.7%96.7%0/343
without_dedup119/19660.7%96.7%0/343
without_tiles119/19660.7%96.7%0/343
without_refinement119/19660.7%96.7%0/343
tiles_nms_only123/19662.8%87.2%0/343
tiles_containment124/19663.3%92.5%0/343
official_faster136/19669.4%66.7%13/343
calibrated_faster132/19667.3%76.3%1/343

Every confirmation image, partial plant, missed plant and failed main-comparison request remains in the denominator. Optional fresh Muse classification was stopped after repeated provider errors; its incomplete cohort is not presented as a complete method.