Hacker News Reader

Comment by Escapado | Hacker News Reader

I wonder if we will see something like this in MS Teams at some point as a lot of my colleagues seem to sit in front of 240p webcams from the 90s with the most atrocious lighting you can imagine.

View on HN

A lot of the resolution issues aren't the camera.

The providers (zoom, webex, MS) have lowered the resolution to lower their bandwidth costs.

> The providers (zoom, webex, MS) have lowered the resolution to lower their bandwidth costs.

Why? Could they NOT do something similar to it - https://blog.google/products/google-plus/saving-you-bandwidt... ?

more ideas - https://www.infoq.com/news/2020/10/nvidia-video-bandwidth/

Because these laptops often have dreadful compute performance as well.

Honestly, it's 90% of the problem. No doubt the crappy bitrates don't help, but it's mostly the camera. Even a fairly low bitrate 480p stream with a good camera looks great.

I've recently started using an old iPhone as a webcam and the differences are startling. Especially in bad light.

Though, it is frustrating seeing how much the bitrate ruins it. Comparing the 1080p output from the iPhone to what you see on zoom etc is depressing. But the camera alone can vary from abysmal to very good. Can't get excellent because of the bitrates though.

I intentionally use a playstation eye webcam...

I'm afraid there's little any algorithm can do if the video you feed it with is that garbage quality, unless it just rebuilds it from scratch using pose estimation etc. but at this point it's just a brand new video.

>there's little any algorithm can do if the video you feed it with is that garbage quality

Have you seen the RTX demo vids? It does seem to defy garbage in garbage out somehow...

When you artificially upscale faces it’s easy to see when it guessed wrong what your face looks like.

I could see live video chat benefiting from you taking a couple close up pictures at the beginning of each chat. Imagine face unlock taking images of your face to later use to deepfake you onto yourself.

ML can do much better than that. It can look at noisy, low resolution video and reconstruct what it would look like on a good camera. Not really an algorithm in the traditional sense.

Reply on news.ycombinator.com