◉

VIDEO ANALYSIS

Fruit Counter

Tests

Find missed fruit, repeated counts and unstable tracks. Change one setting at a time on the same video.

How to choose a useful test

Start with a short, representative clip containing small fruit, clusters and leaves hiding fruit. Use the same clip and manual count across comparisons. Confirm improvements on a second clip before adopting a setting.

More detections do not necessarily mean better accuracy. Review each counted ID against a distinct real fruit. Duplicate IDs for one fruit are false positives. Precision, recall and F1 here evaluate unique-fruit counts after tracking; they are not frame-level mAP. Short tracks are a diagnostic, not proof of an error.

If all settings miss the same fruit, use Extract training frames on the Analysis page to build labeled training and validation data, including difficult lighting and small or occluded fruit. Keep validation videos separate from training.

01

Detection comparisons

Loading comparison settings…

A reference count enables count-error reporting. Precision, recall and F1 require manual review after each attempt.

Settings held constant

Active model, full video, selected classes and the fixed settings above. BoT-SORT, a 30-frame lost-track buffer and counting after 1 observation are used unless that setting is being compared. Other detector and tracker settings use the installed Ultralytics defaults. Each attempt starts with fresh tracking state.

↳

Attempts

Each attempt retains its settings, annotated video and manual review.