Available Metrics

Overview

Arize supports three main computer vision model types, each with specific metrics tailored to their unique characteristics:

1. Object Detection - Detecting and localizing objects in images

2. Image Classification - Classifying images into categories

3. Image Segmentation - Pixel-level classification (Semantic and Instance Segmentation)

Object Detection Metrics

Object Detection models in Arize are designed to detect and localize multiple objects within images using bounding boxes.

Supported Metrics

Primary Metric

• Accuracy - Multi-value accuracy metric that compares predicted bounding box labels with actual bounding box labels

Data Requirements

Object Detection models require the following data fields:

Prediction Data:

• prediction_object_detection_label - List of predicted object labels

• prediction_object_detection_score - Confidence scores for each prediction

• prediction_object_detection_coordinates - Bounding box coordinates

Actual Data:

• actual_object_detection_label - List of ground truth object labels

• actual_object_detection_coordinates - Ground truth bounding box coordinates

Image Classification Metrics

Image Classification models classify entire images into predefined categories. These models support comprehensive multi-class classification metrics.

Supported Metrics

Core Classification Metrics

• Accuracy - Overall classification accuracy

• Precision - Per-class and averaged precision metrics

• Recall - Per-class and averaged recall metrics

• F1 Score - Harmonic mean of precision and recall

• Sensitivity - True positive rate

• Specificity - True negative rate

• False Positive Rate - Rate of incorrect positive predictions

• False Negative Rate - Rate of incorrect negative predictions

• False Negative Density - Density of missed predictions

Multi-Class Specific Metrics

• Multi-Class Precision - Precision calculated per class (requires positive class specification)

• Multi-Class Recall - Recall calculated per class (requires positive class specification)

• Micro-Averaged Precision - Precision averaged across all classes

• Macro-Averaged Precision - Precision averaged across all classes with equal weight

• Micro-Averaged Recall - Recall averaged across all classes

• Macro-Averaged Recall - Recall averaged across all classes with equal weight

Additional Metrics

• AUC - Area Under the ROC Curve

• PR-AUC - Area Under the Precision-Recall Curve

• Log Loss - Cross-entropy loss for probabilistic predictions

• Calibration - Model calibration quality

• Cardinality - Number of unique classes

Data Requirements

Prediction Data:

• prediction_labels - Predicted class labels

• prediction_scores - Confidence scores (optional)

Actual Data:

• actual_labels - Ground truth class labels

Image Segmentation Metrics

Arize supports two types of image segmentation: Semantic Segmentation and Instance Segmentation.

Semantic Segmentation

Semantic segmentation assigns a class label to every pixel in an image.

Supported Metrics

• Accuracy - Multi-value accuracy metric comparing predicted vs actual polygon labels

Data Requirements

Prediction Data:

• prediction_semantic_segmentation_polygon_labels - Predicted segmentation labels

• prediction_semantic_segmentation_polygon_coordinates - Polygon coordinates

Actual Data:

• actual_semantic_segmentation_polygon_labels - Ground truth segmentation labels

• actual_semantic_segmentation_polygon_coordinates - Ground truth polygon coordinates

Instance Segmentation

Instance segmentation identifies and segments individual object instances, combining object detection with segmentation.

Supported Metrics

• Accuracy - Multi-value accuracy metric comparing predicted vs actual polygon labels

Data Requirements

Prediction Data:

• prediction_instance_segmentation_polygon_labels - Predicted instance labels

• prediction_instance_segmentation_polygon_coordinates - Polygon coordinates

• prediction_instance_segmentation_polygon_scores - Confidence scores

• prediction_instance_segmentation_box_coordinates - Bounding box coordinates

Actual Data:

• actual_instance_segmentation_polygon_labels - Ground truth instance labels

• actual_instance_segmentation_polygon_coordinates - Ground truth polygon coordinates

• actual_instance_segmentation_box_coordinates - Ground truth bounding box coordinates