Re:Be My Eyes’ AI assistant starts rolling out
by mattnewton on Wednesday September 27, 2023 @12:49PM
(#37677453)
I’ve actually been fairly impressed with how far monocular depth estimation has come in a few years. I think that it should not be relied upon in self driving cars because they need closer to 100% accuracy and also almost no latency, and the SoTA ones are too slow to be run every frame currently. Not only that but it’s a life-or-death situation where cutting on accuracy % to save BoM costs seems ridiculous to me.
But in a higher-acceptable-latency, lower-risk environment like this, I am actually quite bullish on camera alone methods. Video understanding has come a long way.
Videos may not suffice. Videos are 2d, with 3d aspects being inferred from that 2d data, which is an issue for autonomous driving based on cameras. A proper model for AI training would be 3d scans rather than videos. The best data set would be a combination of video and 3d scanning. Self-driving cars which might combine video with radar/laser scanning may one day provide such a data set.
There is talk of a 3d version of Google streetview, one using a pair of cameras to allow true VR viewing. That might also be good training data as it will capture, in 3d, may street scenes as they unfold.