False positive and false negative rate gaps
Compare false positive and false negative rate gaps across groups.
Testing methods
72FGSM
Test an image model with a single gradient-based adversarial perturbation.
File integrity
Confirm that every image file opens and decodes completely.
Flips
Flip images horizontally or vertically, with their annotations.
Fog
Apply simulated fog to images and compare the model's results with the baseline.
Gamma
Adjust image gamma and compare results with the baseline.
Gaussian noise
Add Gaussian noise to images and compare results with the baseline.
Glare
Add glare to images and compare results with the baseline.
Grad-CAM
Use model feature maps to attribute an image prediction to regions of the input.
Group-wise gaps
Compare group-level classification, regression or oriented-box results.
Image formats
Check that each file's extension matches its decoded image format.
Image sizes
Find small images and image sizes that differ from the rest of the dataset.
Integrated Gradients
Attribute an image prediction by integrating gradients from a reference input.
Intersectional worst-group
Identify the lowest-performing intersection of group attributes.
IoU variants
Compare predicted and reference boxes using IoU, GIoU, DIoU and CIoU.
JPEG compression
Apply JPEG compression and compare results with the baseline.
LIME
Approximate a tabular prediction locally with an interpretable model.
Macro precision, recall and F1
Compare macro-averaged precision, recall and F1 for classification.
MAE
Measure the mean absolute difference between predicted and reference values.
mAP
Measure average precision for oriented bounding boxes.
mAP 50–95
Measure detection average precision across IoU thresholds from 0.50 to 0.95.
MAPE
Measure mean absolute error relative to non-zero reference values.
mAR
Measure detection recall with limits of 1, 10 and 100 detections per image.
Mask mAP
Measure instance-mask average precision at the supported IoU thresholds.
25–48 of 72 results