> Aren't AI models already better at image recognition than humans?

On some benchmarks, AI models are better at very well defined tasks like image classification (“label this image from a set of 8 labels you’ve seen before”) or object detection (“draw a box around all instances of class X in this image, where X is a very narrowly defined class”) They’re not even close to being able to understand unscene examples and parse out their meaning in a larger context the way humans can. (“Recognize that this object in the road is a human riding some sort of bizarre unicycle he welded himself, then predict how he’s likely to move given the structure of his custom unicycle thing”)

The bottleneck in AVs isn’t “perception” in the sense of image classification and object detection, it’s deeper scene understanding and abstract reasoning.

In that case, LIDAR won't give that deeper scene understanding and abstract reasoning, right? It will be some extra input data to the ML model.

The sort of abstract reasoning I’m talking about is beyond the capabilities of any ML model that will run onboard a car - the “abstract reasoning” problem in AVs right now is solved primarily through HD mapping and remote assistance.

LIDAR is useful for a ton of other reasons - ground truth depth, high visibility at night, great for localization - and can detect if something’s an obstacle or not without having seen it before (but false positives can be an issue).

Broadly speaking, people who focus on Camera vs LiDAR are missing the mark and don’t understand that the real difference between big AV players and consumer cars is HD maps and remote support.

LIDAR data contains highly accurate depth information, at which point you don't need the abstract reasoning nearly so much. You can at least do basic object tracking and collision prevention without it.

I've been out of it for a while, but my understanding is that humans are still better at video recognition and extracting high-level information from moving videos, while the state of the art CV still tends to focus on rapid frame-by-frame static classification, with some semblance of persistence and motion strung together at the end.

No, not with edge cases detection. In the Lidar vs Camera context - Camera depth models have a lot of room to improve

There are certain tasks where the models can do better than certain kinds of human effort (e.g. can do better than a person who is tired from doing a thousand of these in a row, is paying very little attention, is spending very little time, and doesn't really care), but that hasn't translated into actually doing image recognition better than humans.

LOL no, not even remotely close. ML image classifiers will decide that an otter is a bicycle with some invisible (to humans) noise injected into the chroma. And that's state-of-the-art classifiers, which Tesla does not possess.

I am asking mostly because the image captchas are starting to be really tough for humans, but ML can easily detect cars, boats and road signs.