Saliency Analysis with XAITK-Saliency#
While NRTK focuses on realistic image perturbations to reveal model degradation at the evaluation level, it is just as important to investigate why this performance degradation may be occuring. A workflow combining NRTK with XAITK-Saliency enables this analysis by generating saliency maps which reveal how model focus changes as perturbations are varied.
The following notebooks showcase this workflow for both image classification and object detection. After completing these tutorials, you’ll be able to:
Apply systematic perturbations and generate interpretable saliency maps.
Recognize robust vs. sensitive model behaviors as you test.
Quantify saliency changes and understand what they indicate.
Make informed decisions about model deployment and improvement.
See also
To explore the underlying perturber APIs used below (Gaussian blur and salt/pepper noise), see the Photometric Perturbers how-to guide.
Combine perturbations with XAITK saliency maps to understand classification model behavior under degradation.
Extend the saliency workflow to object detection, visualizing shifts in detector attention and bounding boxes.