Neural Networks
- Thread starter lunitiks
- Start date
Urgh, 6 Elon months...
I'm curious about how the neural network handle car detection when a car is detected by sides camera and then transition to front camera ? it's a hard problem because of different camera orientation...
They can solve that maybe by re-creating an upper view of the "environment" ? (pure speculation)
How they feed the NN with image buffer ? (Maybe a huge picture that is an consolidation of all camera pictures and by appling some transformation)
They can solve that maybe by re-creating an upper view of the "environment" ? (pure speculation)
How they feed the NN with image buffer ? (Maybe a huge picture that is an consolidation of all camera pictures and by appling some transformation)
Assuming that the NN can detect objects in any orientation (head on, tailing, crossing left/right). Then each camera will detecting the transitioning vehicle. Combine that with known camera positions/ field of view, and object mapping should pass from one to the other without issue.
In this senario, each camera runs its NN separately and only the detected objects are combined at a higher level.
It likely depends on when nvidia contract expires so it is probably sooner than 6 months... unless they are still debugging the silicon
Could be. We also don't know what level the 360 display is pulled from. Functionally, the path algorithm needs to have filtering and blending, but does not need to create a 'good for humans' version/ output. So it may be the 360 view has a different blending algorithm (or none at all).
I need the shorts to help me out ....when they feel the squeeze they will cough up enough to pay for my upgrade to FSD.
I think Elon is talking about the existing chip in that quote, not the new soon to be released one. Elon's 400% lines up with @jimmy_d's 5x the size.
Not quite. Great analysis by Jimmy. I mean top notch. But I was going to come in here gunslinger to burst everyone's bubbles. But elon sorta did it for me.
You see it's not the size of the model that matters, it's the accuracy and efficiency!
Tesla v9 reminds me of a brutefor e approach. In comparison eyeq3 models ran on 0.25 TFLOPS. It's completely unheard of. Today phones come with 3tflop ASIC chips easily (for ex the pixel2 and 3).
The eyeq4 is only 2.5 TFLOP and yet handles all 8 cameras, plus radar and lidar and still have loads of headroom.
It's simply astonishing the NN architecture mobileye were able to create and the efficiency it runs at.
Amon has been very vocal that his approach is different than the bruteforce approach of the industry requiring 50tflops+ chips.
Although he will gladly comply and provide and market them more powerful chips.
You never did explain how Mobileye was able to see a pedestrian kneeling on the floor behind a vehicle... is this what you mean by accuracy?
Don't underestimate what sufficiently advanced NNs could infer, it seeems like magic! They can even be used to drive a car from coast to coast, when run on a Z80!
@J1mbo that is simply a misdetection, false positive that so happens that frames later a pedestrian was near by. If you actually watch in slomo you would see.
EDIT: infact as you can see, parts of the pedestrian is visible. the NN saw it which probably influenced the initial detection
EDIT: infact as you can see, parts of the pedestrian is visible. the NN saw it which probably influenced the initial detection
Last edited:
Don't underestimate what sufficiently advanced NNs could infer, it seeems like magic! They can even be used to drive a car from coast to coast, when run on a Z80!
One of these?
(Z from 80's)Remembering the Soon-to-be-Classic Camaro IROC-Z
(disclaimer: we had a TRS-80)
Joe F
Disruption is hard.
Perhaps, however, if you read the article Lambert wrote, the quoted section of Jimmy's post talked about the increased amount of training data compared to V8, and the increase in weights by a factor of 5. To that section, Lambert quoted Elon's tweet, stating a 400% increase in "useful ops/sec due to enabling integrated GPU & better use of discrete GPU."
This is a disconnect with the block of text Lambert attached Elon's response to.
Joe F
Disruption is hard.
I had an Apple 2+, and bought a Z80 CP/M daughter card for it. Fun times. Wish I had the money I paid back then for all the stuff I paid for computer H/W. Worse was for a memory upgrade for a Gateway after the resin plant fire in Japan in 1993. Oh, and the price a local store wanted for a 5" floppy disk media? Criminal.One of these?
Remembering the Soon-to-be-Classic Camaro IROC-Z
(disclaimer: we had a TRS-80)
jimmy_d
Deep Learning Dork
Don't underestimate what sufficiently advanced NNs could infer, it seeems like magic! They can even be used to drive a car from coast to coast, when run on a Z80!
Hahaha. I resemble that remark.
Similar threads
