GH-ESD: Grounded Hypothesis-Driven Error Slice Discovery for Instance-Level Vision Tasks
Apple Research Blog
Read full postResearchers introduced GH-ESD, a new method for discovering error slices in instance-level vision tasks like object detection and segmentation by generating and verifying grounded hypotheses using large language and vision-language models. They also created the GESD benchmark dataset to evaluate such methods, showing GH-ESD outperforms existing baselines and aids in interpretable model improvements.




