Voxel51 Model Evaluation

From aggregate performance metrics to sample-level diagnostics, Voxel51 allows you to diagnose failure modes and edge cases preventing your models from reaching optimal performance in production.

Try Voxel51 Book a demo

Voxel51 dashboard comparing Model A and Model B with a performance radar chart, prediction statistics, confusion matrices, and distribution charts.

Model Evaluation

Instantly correlate metrics with data samples

Quickly move between aggregate performance metrics and specific data samples to identify exactly what’s driving model failures or successes.

Explore model evaluation

Built-in evaluation methods

Run analyses on regression, classification, detection, polygon, instance, and semantic segmentation tasks. Access standard aggregate metrics like classification reports, confusion matrices, and precision-recall curves directly within Voxel51.

Voxel51 confusion matrices and a radar chart comparing YOLOv8 and R-CNN on mAP, precision, recall, F1 score, and IoU.

Fine-grained, sample-level insights

Go beyond aggregate metrics by capturing detailed statistics like accuracy and false-positive counts at the sample level to reveal annotation errors, training gaps, or data biases.

Overhead view of runners on a track, each in a bounding box labeled runner, next to a Voxel51 panel showing one prediction's label, confidence, and false positive result.

Product Demo

See model evaluation in action

Quickly move between aggregate performance metrics and specific data samples to identify exactly what’s driving model failures or successes.

Model vs. reality

Compare model predictions against ground truth labels directly on your images. Quickly identify where your model excels — and exactly where it needs improvement.

Explore model predictions

$Voxel51 detecting a clavicle fracture on a shoulder X-ray, with predicted and ground truth bounding boxes shown for model evaluation.$

Compare models with scenario analysis

Benchmark models across key metrics and granular data slices to pinpoint performance gaps, identify edge-case failures, and guide targeted model improvements.

Learn more

Voxel51 evaluation summary table comparing two models on IoU, precision, recall, and F1-score over a fruit detection dataset.