Real-time analogue gauge transcription on mobile phone

The objective of this paper is to automatically read any circular single pointer analogue gauge in real-time on mobile phone. We make the following contributions: (i) we show how to efficiently and accurately read gauges on mobile phones using a convolutional neural network (CNN) system which accepts both a high and low resolution gauge image; (ii) we introduce a large synthetic image dataset (far superior in size to prior works) with ground truth gauge readings, pointer layout and scale face homographies that is suitable for training a CNN for real world application; (iii) we also release a new real world analogue gauge dataset (larger meter variation than any previous) with annotation suitable for testing three different types of tasks and finally (iv) we beat state of the art performance for gauge reading on this dataset and an existing public dataset in multiple metrics by large margins, notably with pointer angle error less than 1 degree. Our method is fast and lightweight and runs up to 25fps on mobile devices.