Hacker News Reader

It is done on the CPU instead of the GPU. GPU would seem like the natural choice for a convolution heavy model but was not used here for the mentioned reasons.

Thanks, it makes sense

Some work needs to happen locally to show you a preview of what you're going to transmit, as it should for most video related work.

If the segmentation is done server-side, then you need to sync it to the sender and reflect that quickly in the preview. It's probably not a great experience, at least for a launch.

Probably increased latency due to additional round trips to the server. Perhaps image streams are sent between clients directly, but I'm not sure.