A Large-Scale Synthetic Benchmark for Robust Analog Gauge Reading

Analog gauges are fundamental for data acquisition in numerous industrial sectors, yet the automated reading of these devices is significantly hampered by a scarcity of large-scale, comprehensively annotated datasets. This data bottleneck impedes the development of robust and accurate computer vision algorithms. Here we introduce SyncG, a large-scale synthetic benchmark created to address this critical gap. We developed a generative framework using Blender and Python to produce 20,000 high-resolution, photorealistic gauge images across 145 diverse and challenging environmental settings. Our method allows for precise parametric control over both measurement attributes and visual appearances, ensuring a high degree of diversity and realism. Crucially, we present a fully automated annotation pipeline that generates detailed and accurate ground-truth data for a range of tasks, including object detection, keypoint localization, semantic segmentation, and optical character recognition. By providing this extensive and meticulously annotated benchmark, SyncG facilitates the training and evaluation of sophisticated gauge-reading models and supports the exploration of broader computer vision challenges, such as the interpretation of clock-like graphical representations by multimodal large models.