Is it technologically feasible to identify and recognize gas prices from Google Street View? I just had an idea for an economics research project and would like to know your thoughts. Is this feasible using something like OpenCV? Please forgive my ignorance in terms of the current state of machine-vision technology. I am referring to the standard gas price posts like this: 
 A: Yes, it's technically feasible to do this with current methods. For example, Google uses computer vision algorithms and neural nets to parse out addresses from StreetView images. This is used to match addresses to geospatial locations for Google Maps.

Goodfellow et al. (2013). Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks.
Tech Review: How Google Cracked House Number Identification in Street View

But, it would be a large undertaking. You'd have to solve many difficult problems, and would need access to some pretty heavy resources. For example, you'd need the computational power to process the massive data stream, and would have to develop a complex pipeline just to handle basic operations and coordination between many parallel machines. You'd have to figure out how to identify and extract the relevant image patches containing gas prices, which are square-meter sized regions in a stream of images spanning the planet. It would probably help to use known locations of gas stations, but this is a challenging problem even then (for example, consider the many other signs that could be mistaken for gas prices). From a computational perspective, Goodfellow et al. say that this is the most expensive step (and they don't describe how they do it). You'd have to figure out how to map the noisy, distorted, and possibly occluded image patches to the price and type of gas. Google's Street View network might help. For a sense of scale, the most accurate network in their paper has more than 50 million parameters. Starting with an existing, trained network and then modifying/tuning it for the gas price task would probably be the way to go. You'd also need a way to obtain a large set of ground truth labels for training your algorithm (e.g. army of humans). For reference, Goodfellow et al. trained on a publicly available dataset containing 200k labeled examples, and a private dataset containing 10 million labeled examples.
Short answer: it's probably doable but difficult, and well beyond OpenCV territory. There might be better avenues for obtaining the data. Aren't there smartphone apps that give gas prices in realtime?
