A Deep Research Agent for Curating Vision Datasets
10 hours ago
- #Data Curation
- #Computer Vision
- #AI Research
- Labeling Copilot is introduced as the first deep research agent for automated data curation in computer vision.
- It features a central orchestrator agent powered by a large multimodal language model for multi-step reasoning.
- Core capabilities include Calibrated Discovery, Controllable Synthesis, and Consensus Annotation.
- Consensus Annotation excels in object discovery, achieving high annotation accuracy and expanding category coverage.
- Calibrated Discovery is computationally efficient, with an active learning strategy that outperforms alternatives.
- The system is validated on large-scale datasets like COCO and Open Images, demonstrating robust performance.