Tesla certainly thinks so. Read what its chief AI scientist Andrej Karpathy has to say. I remember Dr Karpathy from his widely quoted article The Unreasonable Effectiveness of Recurrent Neural Networks. He got his PhD under CV pioneer Prof Fei-Fei Li.
Alright, it’s time to revisit this discussion with Tesla’s AI Day!
Andrej Karpathy presented some pretty amazing content in the “Tesla Vision” segment of the presentation – using models that predict in “vector space” instead of traditional image space to get more accurate representations of the environment; using variants of recurrent neural networks to “remember” objects over time and solve occlusion problems; using AI instead of traditional heuristics such as A* search algorithms to optimise path planning… and more!
Also interesting to note the use of 4D labelling of the environment instead of traditional labelling of 2D images, and the use of simulation to train models on edge cases and complex scenarios.
Tesla’s vertical integration approach to it’s manufacturing has also permeated its AI division – data labellers are in-house, they wrote their own simulation and data labelling tools, and designed their own FSD and DOJO computers.
Very exciting stuff and I look forward to more cutting edge developments from them!
Just to link it back to the question – is CV sufficient for self driving?
The bottom left image shows the interpretation of a traffic junction using Tesla’s 2017 tech in image space.
The bottom right image shows the traffic junction using the new vector space approach – it’s so clearly defined that I wouldn’t have thought it would be possible with just multiple cameras and no LiDAR!
I can’t answer this question yet, but it certainly looks like Tesla has taken a giant leap forward in this area…
This reply was modified 3 weeks, 5 days ago by leeping-ng.