Evaluate a model against your own ground truth

Model evaluation runs an available prediction engine on annotated images in the saved test split, scores its predictions locally and keeps the result with the scope that produced it. It helps you choose a model or find weak labels before another annotation pass.

Prepare and run an evaluation

  1. Review your test images and apply a saved train/validation/test split. Unannotated images are excluded rather than counted as empty ground truth.
  2. Open the model evaluation panel and select a ready active model. Text-prompt and visual-prompt engines are currently disabled for reproducible benchmarks because they do not expose pinned raw outputs. Task compatibility and device readiness still apply.
  3. Set the confidence threshold, inspect the scope and run evaluation. Progress separates preparation, inference and scoring; you can cancel a run.
  4. Inspect overall and per-label results. Compare only runs with compatible data, label schema, split and settings; a changed scope is not the same benchmark.

Metrics match the project task

TaskReported measurements
Detection and instance segmentationPrecision, recall, F1, mAP@0.50 and mAP@0.50:0.95, including per-label scores; overlap follows the task geometry.
Semantic segmentationPer-class pixel IoU and mean IoU, using the declared semantic task.
ClassificationAccuracy, macro F1 and per-label scores; a confusion matrix for single-label ground truth.
KeypointsPCK: the fraction of visible points within the configured positional tolerance.

A metric implementation does not make every engine available for every task. The engine picker checks workflow compatibility and device readiness. This benchmark uses images; it does not score a native video timeline.

The score includes its evidence

Saved results identify the scored images, pixel and ground-truth identities, task and label schema, test assignments, split configuration and engine settings. Scope changes are detected instead of silently attaching an old score to new data. Exclusions are counted so missing labels cannot look like verified negatives.

See also

Video tutorial

AnnotateIt tutorial