Which tasks can OvalBee evaluate?
Object detection: scoring a single model, comparing up to five, measuring pre-labeling efficiency and tracking an active learning curve. Evaluation for classification, segmentation and tracking is on the way. If your project needs it now, tell us during onboarding.
Can I evaluate a model trained outside OvalBee?
Yes. Evaluation compares a dataset of predictions with a dataset of ground truth, so it does not matter where the model came from. Upload its weights with the Upload model node and run a predict node, or import predictions you already have.
How do I know a comparison is fair?
Pin the test set to a version, so it cannot change between runs. Before ranking, the comparison checks that every model was scored against the same ground truth, classes and thresholds, and Benchmarks warns when the class sets of two runs differ.
How is the winner chosen?
By the highest mAP; a shared top score is reported as a tie. The report also names the leader in AP50, AP75, precision, recall and F1, and the winner is an output of the node, so the next step in the flow can use whichever model won.
What does “mAP gained if fixed” mean?
Every wrong or missing box is sorted into one type: wrong class, loose box, background, missed object or duplicate. The report then recomputes mAP as if one type of error were fixed, which shows where more data or a different setup would pay off most.
Why does a class score zero?
Most often its name differs between the ground truth and the model, for example stop_sign and stop sign. Unmatched names count as missed objects, and the frequently confused classes section of the report shows such pairs with a confusion of 1.0.
How is the confidence threshold chosen?
By default, the report picks the threshold with the best F1. You can set your own threshold, IoU from 0.50 to 0.95 and the maximum number of detections per image, and every number in the report follows.
Can I deploy the model that wins?
Yes. Deploy a YOLO version as an HTTP endpoint on the machine you choose, with an API token on every request and logs in the app. YOLO, D-FINE, RT-DETRv2 and DEIMv2 models also export to ONNX and TensorRT for your own hardware.
Can I share a report?
Yes. Save any report as HTML or print it to PDF. Every run also stays attached to the model version it scored, so the evidence behind a decision does not get lost.
ripeunripeunripe