Evaluation Settings
Look up the controls that affect prediction generation, scoring, and the tutorial Fog comparison.
Prediction Generation controls which outputs and scenario coverage are available. Evaluation selects measurements and criteria from that retained Prediction Set. If new settings require predictions that are not present, generate a new Prediction Set first.
Inputs and inference
The example values below describe the Helmet tutorial. They are not universal defaults for every model or task.
| Control | Purpose | Helmet tutorial value |
|---|---|---|
| Dataset | Imported samples and reference labels | Helmet v2 full dataset |
| Model / API Profile | Source that produces predictions | Helmet YOLO26 detector, a local YOLO model |
| Class labels | Names in model output order; align them with dataset labels | Class 0: helmet |
| Input width / height | Fixed dimensions required by the selected model | 640 × 640 pixels |
| Model runner | Compatible inference implementation | YOLO |
| Dataset split | Which partition supplies samples | Validation |
| Sample scope | All selected-split samples or a configured subset | Random examples: 100 of 195; seed 42 |
| Minimum object confidence | Filters detections below the chosen confidence | 0.25 |
| Duplicate-box overlap | Overlap threshold for suppressing duplicate detections | 0.5 |
Increasing minimum confidence can remove false positives and also remove valid low-confidence detections. Changing it changes the prediction evidence. Keep it fixed when you want to compare only the effect of a robustness transform.
The duplicate-box setting controls inference postprocessing; it does not replace the matching rule used by Detection mAP 50-95.
For supported dataset and model files, use the format table.
Robustness coverage and scoring
| Control | Meaning | Example |
|---|---|---|
| Transform | Modification applied to inputs | Fog |
| Severity | Selected image-transform level; integer 1 through 5 | 3 |
| Seed | Non-negative base seed for stochastic transforms | 0 |
| Performance metric | Measurement compared between clean and transformed inputs | Detection mAP 50-95 |
| Maximum allowed degradation (%) | Acceptance limit from 0 to 100; blank means no criterion | 20 |
A severity label selects a transform recipe. It does not describe a percentage of fog, and the same severity number means different settings for different transforms. The degradation limit changes the decision rule, not the image transformation.
Fog settings by severity
| Severity | Fog strength | Wibble decay |
|---|---|---|
| 1 | 1.5 | 2.0 |
| 2 | 2.0 | 2.0 |
| 3 | 2.5 | 1.7 |
| 4 | 2.5 | 1.5 |
| 5 | 3.0 | 1.4 |
Fog is stochastic. Preserve the base seed, sample identities, input preparation, and runtime version as well as the severity. The result and report retain the resolved parameters used by that run; use those recorded values when comparing or reproducing results.
What requires a new record?
| Change | Next action |
|---|---|
| Model weights, dataset contents, inference thresholds, split, or sample scope | Prepare compatible inputs and generate a new Prediction Set |
| Add a transform or severity absent from the sealed set | Generate a set containing the required coverage |
| Change metrics or criteria with compatible retained evidence | Create a new Evaluation using the sealed set |
| Change report wording or included sections | Edit the report draft and export a new report |
See Metrics, Thresholds, and Robustness for score interpretation and Evaluation Presets for reusing a reviewed configuration.