SliceLens: Fine-Grained Error Slice Discovery for Multi-Instance Vision

This paper addresses the critical challenge of identifying and understanding systematic failures (error slices) in computer vision models, particularly for multi-instance tasks like object detection and segmentation. It highlights the limitations of existing methods, especially their inability to handle complex visual relationships and the lack of suitable benchmarks. The proposed SliceLens framework leverages LLMs and VLMs for hypothesis generation and verification, leading to more interpretable and actionable insights. The introduction of the FeSD benchmark is a significant contribution, providing a more realistic and fine-grained evaluation environment. The paper's focus on improving model robustness and providing actionable insights makes it valuable for researchers and practitioners in computer vision.