It is unclear from the images shown what the ML is actually adding to the enhancement that couldn't be achieved by just loading the image into your favourite photo editing software and fiddling with the brightness. To fix this, it would be a good idea to show images that have been adjusted in this way instead of the black images. Then we can see the difference between non-ML adjustment and ML enhancement.
I thought this too - so I just tried it. It's significantly better than using photoshop. The colours are MUCH better and the edges are also more accurate, you would guess it was a higher resolution image.
well, I just looked at your image and if you run some blur over your "best", it looks a lot better and it's not like there's a huge qualitative difference in actual resolution/visible detail - even when you are comparing your improvement on the posted mini-image with the ML-generated one supposedly from a non-compressed/scaled image. And in high-res you then would see the inevitable artifacts of "one-filter-fits-all-CNNs"...
The results are visually striking but also to be expected if the sensors already captured the necessary data (to the computer) and is obscured (to the human) by the relative (instead of absolute) perception.
What I find innovative here is the concept of thinking about anything that is hard to do as a human (night vision) but for which we can affect the state of (in this case, the light switch) in order to train a model to overcome the difficulty.
Sorry to belabor the obvious but I thought it was worth a description which may motivate other use cases.
A repost of https://news.ycombinator.com/item?id=26487503 which was inexplicably dead, perhaps for linking to Twitter instead of Github. Auto-killing posts doesn't really work very well on HN.
Note that he did not write the paper nor model, nor perform the training. This repo is only downloading a trained model, using a tool to convert it to another format, and using a canned API to plot it and do some inference.
Definitely a high-flying student with a bright future ahead, but this is just stitching together a few API calls.
