Fine-Grained Annotation Through Multi-Model Fusion and Class-Specific Performance Optimization in Object Detection
Liu Hongzhi · 2024
Automatic annotation is essential for large-scale data labeling in object detection, but traditional single-model methods often lack accuracy and reliability. Multi-model systems address this by leveraging the strengths of multiple detectors but face two key challenges: identifying whether predictions from different models refer to the same target and effectively integrating those predictions. This paper introduces a fine-grained annotation method and a distributed system architecture to address these issues. A target matching mechanism inspired by non-maximum suppression (NMS) consolidates predictions across models by evaluating bounding box overlaps. A dynamic decision-making algorithm assigns adaptive voting weights based on classspecific model performance and uncertainty, ensuring precise and robust label assignments. The distributed architecture, combining a storage layer, search engine, and scheduling servers, efficiently handles largescale annotation tasks. This approach highlights the benefits of multi-model integration, adaptive decision-making, and scalable design for reliable automatic annotation in object detection.