Holistic Video Scene Understanding with ViP-DeepLab

[diplomat33]

https://twitter.com/x/status/1387141759164194820

Full blog: Holistic Video Scene Understanding with ViP-DeepLab

 

[ZeApelido]

High quality camera vision perception system is not possible - ask all the machine learning experts on this forum...

 

[Soda Popinski]

Pretty cool. I wonder in the context of Tesla, if the FSD computer is capable of handling both. We've seen Pseudo-Lidar from the Karpathy talk, which must be monocular for the side cameras at least. I'm also unclear if the standard object/image recognition used is considered panoptic segmentation.

 

[diplomat33]

High quality camera vision perception system is not possible - ask all the machine learning experts on this forum...

I don't think anyone is saying that high quality camera vision perception is not possible. I am certainly not saying that. The question is how good can camera vision get? Can it be 99.999999% reliable for all cases? And are there any benefits in having other sensors for redundancy like for cases where camera vision might be less reliable?

 

[EVNow]

I don't think anyone is saying that high quality camera vision perception is not possible. I am certainly not saying that. The question is how good can camera vision get? Can it be 99.999999% reliable for all cases? And are there any benefits in having other sensors for redundancy like for cases where camera vision might be less reliable?

They didn't start using Lidar for redundancy.

 

[J1mbo]

Not massively impressed with the 3D view from that example. The "shadow" cast by the two pedestians is excessive and obliterates objects that we know are there from previous frames (i.e. the blue oncoming car).

Bonus point for making one of the pedestrians look like a character out of The Simpsons