Been thinking about weather for a while. We ought to be able to get the sun’s position on the screen, and by extension whether the image is looking into the sun, or whether it’s day or night time, based on the location, time and orientation of the shot, and identify images that are likely bad ones. It could also give clues about the direction of shadows. I was thinking that encouraging but filtering out night time images (at least by default) would make it much easier to detect street lighting and lit road signs… there’s a wealth of different data available at night.
What would be cooler, is if the app somehow stored local weather conditions in the EXIF metadata. This could give good clues about the sharpness of shadows or whether there’s likely to be rain on the road or windscreen, how likely snow or ice is (I’ve shots here that detect snow in Britain in mid-August - clearly a false detection). Gathering historic weather data would likely be expensive and awkward, but there’s probably lots of ways to gather that data on a phone at the time of capture for free, and it’d be tiny amounts of data compared to pixels - well worth save it in case it’s useful to future detection models, or just having a way to filter out images of crappy days in my tourist town.
Yes the sun’s position can be calculated, no need for that to go in the EXIF metadata. The weather information is generally statistical, 30°C, sunny with a 1% chance of rain would make a grid shape more likely to be the shadow of a fence than a drainage grid, and a white surface unlikely to be snow, while 3°and a 90% chance of rain would make the opposite more likely. Overcast pictures would be better, on average, for painting textures onto 3D surfaces without being soiled by shadows, while pictures taken in sunshine would be better for texture extraction.
Capturing the info would make more stuff possible, so it’s a shame to not do it if it’s free.
Yeah I understand the desire to keep data that’s not directly related to the image away from the image, it’s unhygienic, but I don’t think it’d be easier to do this sort of thing later. If you have the data in each image then you have bulk, historical data, if not then you need to source and license that data. I guess I value pragmatism over technical purity, and have a strong dislike of external dependencies.
Does the Android app upload images from other devices? I was talking about images captured on smartphones that generally have internet access and get their time from the network.
The fundamental problem with the weather is that any data pulled it isn’t necessarily accurate. There are a lot of issues with reporting at weather stations. And often the weather at the moment a mapping photo is generated is different than the nearest weather stations 2km away.
But that’s a known issue. I mean, all information gathered has a fuzzy truth value, GPS coordinates, compass direction, object recognitions and so on are all inaccurate to some degree. The question is whether the information is useful. I think it likely is
)