Sketch explainability plays a crucial role in diverse downstream sketch-related tasks, offering insights beyond network behavior.
Abstract
The paper explores the significance of human strokes in sketch explainability, proposing a lightweight solution for seamless integration with pre-trained models.
Applications include retrieval, generation, assisted drawing, and adversarial attacks on sketches.
Stroke-level attribution maps provide insights into model predictions and behavior across various tasks.
Evaluation metrics include mAP, precision, accuracy, and mean opinion score (MOS) for different applications.
Human studies assess transparency, fairness, and trustworthiness of the model's predictions based on stroke attributions.
Customize Summary
Rewrite with AI
Generate Citations
Translate Source
To Another Language
Generate MindMap
from source content
Visit Source
arxiv.org
What Sketch Explainability Really Means for Downstream Tasks