Magnitude
MAE, RMSE, TRE and surface distances answer “how far off?” and usually retain the physical or target unit.
OPEN SOFTWARE · SCIENTIFIC EVALUATION STUDIO
OpenMetricLab
Upload actual values and predictions and turn them into a clear scientific report. Compare models, inspect failure patterns, and export results for regression, classification, segmentation, and image registration.
RegressionClassificationSegmentationImage registrationBrowser + mobileLocal processingMIT licensed
01LoadTruth + predictions
02InspectScores + diagnostics
03ExportReport + data
START WITH THE TASK
Choose what “actual vs prediction” means in your experiment.
The input and the correct metric family depend on the task. OpenMetricLab only shows metrics that are meaningful for the information you supplied; probability metrics require probabilities, boundary metrics require masks, and landmark error requires landmarks.
Regression
Upload one CSV, choose truth and predictions.
1 · Load2 · Configure3 · Analyze
METRIC FAMILIES, NOT METRIC SHOPPING
Each metric answers a different scientific question.
The interface deliberately prioritizes complementary metric families rather than encouraging users to select whichever single number looks best.
Balance
Balanced accuracy, macro-F1, per-class recall and MCC reveal performance that raw accuracy can hide under class imbalance.
Overlap + boundary
Dice/IoU measure region agreement; HD95 and ASSD measure spatial boundary disagreement. They are not interchangeable.
Similarity + geometry
SSIM/NCC/NMI compare image similarity. Landmark TRE evaluates geometric correspondence when landmarks are available.
BUILT-IN REPRODUCIBLE EXAMPLES
Learn the interface on real public datasets before using your own results.
The examples are fetched from udanish50/OpenMetricLab and are reproducible. Real public datasets are used where a natural truth/prediction task exists; deterministic perturbations are clearly identified where the goal is to demonstrate an image metric rather than claim model performance.
| Task | Example | Source | What it demonstrates |
|---|---|---|---|
| Regression | Diabetes | Bundled public dataset | Two prediction columns, model comparison, residuals and tail error. |
| Classification | Iris, Wine, Breast Cancer, Digits | Bundled public dataset | Binary/multiclass labels and class-probability evaluation. |
| Segmentation | Horse silhouette | Bundled public/standard image | Binary overlap, error overlay and surface-distance metrics with a deterministic perturbed prediction. |
| Segmentation | Three-label geometry | Deterministic fixture | Multiclass per-label reporting and macro aggregation. |
| Registration | Shepp–Logan phantom | Bundled public/standard image | Before/after image similarity, checkerboard review, and landmark TRE. |
Example predictions are demonstration artifacts generated deterministically. They are not presented as state-of-the-art results or as published benchmark claims.
REPORT LIKE A RESEARCHER
A good evaluation explains the score and the failure pattern.
Metric cards are followed by visual diagnostics and an interpretation narrative. Reports can be printed/saved as PDF, exported as JSON, or reduced to a metrics CSV for manuscripts and supplementary material.
Show uncertainty in interpretation
R² can be negative; MAPE can be undefined at zero targets; image similarity is not geometric correctness. The interface surfaces these caveats instead of hiding them.
Keep units visible
RMSE, TRE, HD95 and ASSD are meaningful only with the target or spatial unit. Pixel/voxel spacing inputs are available for image tasks.
Inspect errors, not only means
Residual plots, confusion matrices, per-class tables, mask error overlays and checkerboards show where the model fails.
Do not tune on the test report
Threshold exploration is diagnostic. Threshold selection and hyperparameter choices belong on training/validation data, not the held-out final test set.
SOURCE & REPRODUCIBILITY
Transparent definitions and reproducible calculations.
Metric definitions, edge-case policies, tests, and source code are published with the software. Browser reports follow the documented definitions used by the project so results can be inspected, reproduced, and exported.
Implementation note
Core metric calculations are implemented in the project source and checked with hand-verifiable examples, edge cases, and reproducibility tests. The technical details stay here so the main interface can focus on evaluation.
Metric definitions, assumptions, edge-case behavior, example provenance, and reproducibility notes are documented in the repository alongside the source and tests.