Run several ready-made detectors on the same labeled test set and get an objective, measured winner — no training required.
▶ Open this template in OvalBee
Overview
- Model: Several ready-made detectors, scored side by side — no training involved (the candidates are a parameter; the linked run compares YOLOv8s, EdgeCrafter and DEIM)
- Task: Object detection
- Input: A labeled image set with ground truth, plus the set of candidate models to compare — both supplied as parameters
- Pipeline: pick the labeled set → run every candidate detector on it → score each against the same ground truth → compare head-to-head and pick a winner
- Output: A per-model accuracy comparison (mAP, AP50, AP75, F1), a per-class breakdown where the data supports it, a declared winner, and side-by-side prediction previews on the same images
When you need it
This is the template for the moment you have several candidate detectors — different architectures, different versions, different checkpoints — and none of them need any training: they're all ready to run, and the only question left is which one to deploy. A hunch, or three separately run benchmarks with slightly different settings, isn't a defensible answer, so this template puts every candidate through the identical test — same images, same ground truth, same scoring code, one run — so the winner it declares is a measured result, not a guess dressed up as one.
What you need
A labeled image set with real ground truth to score against, and the set of candidate models you want to compare — as few as two, or several at once. Neither is trained here; every candidate arrives ready to run, and the dataset and the candidate list are both parameters. If your labeled set isn't in OvalBee yet, bring it in with one of the import templates first.
How it works
| Step | What happens |
|---|---|
| Dataset | Pick the labeled image set to test, ground truth included |
| Predict | Every candidate detector runs on the same images |
| Score | Each model is measured against the same ground truth |
| Compare | The candidates go head-to-head, per class where the data supports it, and a winner is picked |
What a run gives you
A side-by-side accuracy table across every candidate — mAP with AP50, AP75 and F1 — plus a per-class breakdown wherever the test set has enough ground-truth examples of that class to back it up, a declared winner, and prediction previews so you can see ground truth against every candidate's boxes on the same images, not just read their scores.
Because every candidate runs through the same OvalBee comparison, the verdict is one you can defend: the same held-out images and the same scoring code for all of them, not separate evaluations where a slightly different setting could quietly tilt the numbers. Add a new candidate later and OvalBee runs the identical test again, so the next architecture you're weighing gets the same credible verdict without rebuilding the comparison by hand.
Related templates
- Detector training — if none of your ready-made candidates are good enough, train one on your own labeled classes instead
- Detector evaluation — the single-candidate version of this same scoring step, for accepting or rejecting one model rather than choosing between several
- Two trained detectors is a sibling template in this library — the same head-to-head comparison for models you trained yourself rather than ready-made ones — not yet published here