Reproduce “DART: An Automated End-to-End Object Detection Pipeline with Data Diversification, Open-Vocabulary Bounding Box Annotation, Pseudo-Label Review, and Model Training” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Chen Xin, Andreas Hartel, Enkelejda Kasneci

Researchers at Technical University of Munich and Liebherr-Electronics and Drives developed DART, an automated, end-to-end pipeline for object detection that eliminates manual annotation. The system integrates generative AI, open-vocabulary detection, and large multimodal models to achieve a twelve-fold increase in average precision for industrial object detection tasks, improving YOLOv8n's AP from 0.064 to 0.832.

Key insights

  • An automated, synergistic integration of generative AI, open-vocabulary detection, and LMMs can eliminate manual annotation burdens and data scarcity for specialized object detection.
  • LMM-based review is critical for curating high-fidelity synthetic data and accurate pseudo-labels, preventing negative impacts on downstream model performance.
  • Combining diversified synthetic data (e.g., a 3:1 ratio with original data) significantly enhances model robustness and generalization, leading to substantial accuracy gains.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.