Reproduce “Evaluating the Performance of Open-Vocabulary Object Detection in Low-quality Image” 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

Po-Chih Wu

This study evaluates the performance of open-vocabulary object detection models under various image degradations by introducing a new 'Low-quality image dataset'. It reveals that while models like OWLv2-L/14 exhibit greater robustness, all state-of-the-art models experience a significant decline in detection accuracy when faced with severe image quality degradation.

Key insights

  • Open-vocabulary object detection models demonstrate a general vulnerability to severe image degradation, showing a sharp decline in mAP scores when degradation levels are high.
  • Minor image degradations typically have a negligible impact on performance, but a threshold exists beyond which severe degradation leads to significant drops across all evaluated models.
  • OWLv2 models, particularly OWLv2-L/14, consistently exhibited superior stability and higher mAP scores across all degradation types and severity levels, indicating greater inherent robustness.

Summary by alphaXiv.

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