4 ms·
There is perhaps an interesting machine learning application here. There are higher resolution images available at lower frequencies and low resolution images
by manquer 5y ago
There is perhaps an interesting machine learning application here.
There are higher resolution images available at lower frequencies and low resolution images possible with high frequency.
Would it not possible to "zoom and enhance" a low resolution image to higher resolution one using historical high res data of the location and learning from object types and classification?
- agnokapathetic 5y agoYes: https://medium.com/the-downlinq/super-resolution-and-object-detection-a-love-story-part-1a-2f1e962e99dd https://medium.com/the-downlinq/super-resolution-and-object-...
- gruez 5y agoYou're basically letting an AI to hallucinate details, which looks good in most cases but in edge cases may cause unintended results: https://news.ycombinator.com/item?id=24196650 https://news.ycombinator.com/item?id=24196650
- vinhboy 5y agoYo. That was crazy. So the whole "zoom and enhance" thing they do on TV is no longer science fiction. That meme is dead judging by what I am seeing in that thread.
- fwip 5y agoWe can get computers to make things up, sure. But usually when they "zoom and enhance" on TV it's because they want to see what's really there.
- gruez 5y agoYeah, if you used this sort of upscale AI on a pixelated picture of a suspect, you'll just end up arresting some guy who's face looks most common in the training dataset.
- jjgreen 5y agoKind-of anticipated by Antonioni in Blow Up.
- colechristensen 5y agoIt is absolutely possible and there are plenty of ML and non-ML solutions for this or similar problems. A set of useful keywords would be "image sensor fusion". Of course there would be limitations and most of the image updates would be boring (all the buildings and streets are still there) or lacking enough information to retrieve anything (new set of cars parked on the street) so you'd end up with highlights and uncertainties for changes overlaid on top of high confidence existing infra.
- dragonwriter 5y ago> Would it not possible to "zoom and enhance" a low resolution image to higher resolution one using historical high res data of the location and learning from object types and classification? Sure. Probably provide pretty convincing high-detail images. For the cases of most interest—i.e., when the new activity is unusual and unexpected—it will often be detailed, convincing, and completely wrong, though.
- manquer 5y agoThere are definitely applications where those details don't matter , that is why low res services do exist after all. It would be nice UX enchantment if you will to what is a low res product.
- bombcar 5y agoThe more sources of information you have the better you can collect and correlate it - especially if you have various low-resolutions taken at different times but still able to be lined up.