Reproduce “Under pressure: learning-based analog gauge reading in the wild” 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
Maurits Reitsma, Julian Keller, Kenneth Blomqvist, Roland Siegwart
Researchers at ETH Zurich developed a comprehensive learning-based framework to automate analog gauge reading in real-world industrial settings. This modular pipeline effectively handles gauge variability and perspective distortion, achieving approximately 2% relative reading error across diverse datasets without requiring prior knowledge of gauge specifications.
Key insights
- The modular pipeline design allows for independent optimization of each stage, robust failure detection, and component-wise upgrades, making the system adaptable and diagnosable.
- Utilizing keypoint detection with a DINOv2 backbone to identify major scale notches enables robust estimation of scale geometry, even with partial occlusions or varying image quality.
- Fitting an ellipse to the gauge scale and applying perspective correction based on this model significantly improves OCR accuracy, particularly for gauges viewed at oblique angles.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.