Apple researchers developed GH-ESD to identify "error slices" in instance-level vision tasks like object detection. Unlike previous methods that rely on representation clusters, this approach targets spatially grounded visual patterns and contextual relations. It allows developers to pinpoint specific semantic failures. This improves how teams debug robustness in complex segmentation models.