Advanced AI today can really have some trippy versions of reality if you let them go wild. To their credit, it helps with overall image recognition though. It gives our digital witnesses a chance to show just how much they could learn and perceive in our world.
You know how much more AI could prove this? Why by recreating realistic scenes from scratch! Perhaps, just like what this latest AI research by Stanford and Intel are currently working on.
The Not-so-Realistic Haze
The image you are seeing right now isn’t a snapshot taken by a really low-end camera. It’s completely artificial, at least, in the way that it is designed by Vladlen Koltun and Qifeng Chen’s latest AI creation. This collaborative project between Stanford and Intel shows just how much AI could perceive a ‘natural setting’ when given enough data.
According to the demonstration video, the AI achieves this level of instructional understanding by first analyzing more than 5,000 photos from German’s urban areas. A basic instruction set is then given as input, which then the AI uses as a base design layout for a ‘scenery’ that it will create from the data it has collected.
For example, different colors are designated with different objects that can be used for the image. The AI fills up all of the spaces within the crude color-coded image, appropriately allocating the right type of object, with the (more or less) correct positioning, contrast and/or lighting within the image space. The resulting image is as you see above. Squint a bit more and you’ll easily notice it’s less-refined qualities. However, as a general (blurred) low-res image, it does look convincingly, even a bit scarily, real enough.
Rendering in the Matrix
Okay, so it might not be realistic enough yet, but the technology does show a lot of promise. If an AI could design almost convincingly real scenes like this even with such crude resolutions, we could simply extrapolate in the future the level of detail it could do as it learns more and more.
In fact, there might even be certain merit in using the technology as it is right now. Mobile VR, in comparison to high-end desktop VR, is significantly less detailed. Realistic enough, but still kinda blurry. Most mobile VR apps are designed to give the basic experience of 360-degree exploration without the gimmick of ultra-high texture detail, or top-end frame rates.
Perhaps in the very near future, this type of AI can be used to quickly design traversable areas for mobile VR apps. The look and feel should be realistic enough, but is still within the performance confines of your average smartphone or tablet.
Source: Intel VCL