twitter-research/image-crop-analysis

Implements saliency-based image crop prediction with demographic bias analysis across intersectional population groups, using saliency maps to evaluate fairness metrics. The analysis pipeline includes data preparation, statistical testing, and visualization notebooks that measure disparate impact and representational harms. Runnable via Jupyter notebooks, Docker, or Google Colab with scikit-learn and scikit-image for computer vision processing.

253 stars. No commits in the last 6 months.