Reproduce “Semi-Supervised Object Detection: A Survey on Progress from CNN to Transformer” 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

Tahira Shehzadi, Ifza, Didier Stricker, Muhammad Zeshan Afzal

Researchers from the Technical University of Kaiserslautern and DFKI provide a comprehensive survey of Semi-Supervised Object Detection (SSOD), charting its evolution from CNN-based architectures to modern Transformer-based models. The work systematically analyzes 27 cutting-edge methodologies, demonstrating how Transformer-based SSOD models like Sparse Semi-DETR achieve superior performance, for instance, reaching 44.30 AP on COCO-Partial 10% compared to typical CNN-based models ranging from 30-37 AP.

Key insights

  • Pseudo-labeling and consistency regularization are identified as the two most fundamental and prevalent strategies in modern SSOD.
  • Transformer-based architectures offer superior capabilities in capturing long-range dependencies and handling complex spatial arrangements, leading to performance gains in SSOD.
  • Effective data augmentation, including weak, strong, and hybrid techniques, is crucial for creating diverse views of unlabeled data and improving model robustness.

Summary by alphaXiv.

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