-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
There will be NO HW4 upgrade for HW3 owners
- Thread starter macmee
- Start date
Most FSDb problems I encounter seem more like planning issues than perception. Given that Tesla has indicated that they are in process of converting to NNs for all planning and control, it would seem likely to have a few planning issues.
stopcrazypp
Well-Known Member
From what I can find the difference in FOV is so tiny that it won't make any significant difference.
This image is the more accurate one from below thread:
A new camera position would be what would make a difference and the Y HW4 has that depopulated so it's not an option in the current boards.
2 new cameras under the windshield- one angled towards the left, and another towards the right in an X pattern would be best solution.From what I can find the difference in FOV is so tiny that it won't make any significant difference.
This image is the more accurate one from below thread:
A new camera position would be what would make a difference and the Y HW4 has that depopulated so it's not an option in the current boards.
Hank42
Legacy Supporter
And would be trivial to retrofit into HW3 cars as well.
How do we know Tesla isn't cropping the view shown on screen and in the dashcam footage? I'm not sure how much weight we should really put on the example shared. They may not want to make it obvious that HW4 repeater cams are significantly better (they're still trying to sell HW3 cars).
stopcrazypp
Well-Known Member
From other thread, we know because they appear to be doing 2x2 binning on the repeater cams, judging from the 5.4MP sensors comparing repeater and front camera resolution:
"1448 × 938 (1.4MP) resolution for repeaters and 2896 × 1876 (5.4MP) for the front"
First Tesla vehicle delivered with HW4
These specs match the Sony IMX490, which was said to be the selected camera supplier for HW4:
Tesla's FSD hardware 4.0 to use cameras with LED flicker mitigation
Image sensor gives 2896 x 1876 full resolution, so Tesla is using the full resolution already:
https://www.sony-semicon.com/files/62/pdf/p-15_IMX490.pdf
If they were doing cropping to normalize to HW3 they would instead try to match the 1280x960 (1.2MP) of the older cameras.
Last edited:
KArnold
Active Member
stopcrazypp
Well-Known Member
He simply understood incorrectly.
HW4 has upgrades to the FSD board, and the MCU infotainment board integrates the separate GPU daughterboard, but the FSD and MCU are still separate boards.
New Tesla HW4 computer analysis confirms it won't fit older models but will save Tesla millions in AMD chip integration
@stopcrazypp From reading that other thread I'm still not quite sure I'm convinced that they're not cropping. Obviously they could be doing 2x2 binning too but the evidence in support of either seems really sparse.
I guess the main evidence is that they choose to use a 1448 × 938 resolution (vs matching HW3 1280x960)? That does seem like a strange choice, although not one that's makes cropping impossible.
I might be missing part of your reasoning too, you've done quite a bit of thinking about this it seems!
To me the ultrawide repeater cam just seemed like such a nice and elegant solution to the problem of side visibility at T intersections with poor visibility. That seemed like a big weakness of the current camera positions. Perhaps I just really want them to do that and it's clouding my judgement of the actual evidence at hand.
I guess the main evidence is that they choose to use a 1448 × 938 resolution (vs matching HW3 1280x960)? That does seem like a strange choice, although not one that's makes cropping impossible.
I might be missing part of your reasoning too, you've done quite a bit of thinking about this it seems!
To me the ultrawide repeater cam just seemed like such a nice and elegant solution to the problem of side visibility at T intersections with poor visibility. That seemed like a big weakness of the current camera positions. Perhaps I just really want them to do that and it's clouding my judgement of the actual evidence at hand.
stopcrazypp
Well-Known Member
Maybe you missed my edit, the full sensor resolution of the Sony IMX490 Tesla is using is exactly 2896 x 1876, a native 2x2 binning mode would be 1448 x 938. The spec sheet I linked of the sensor doesn't have the binning modes listed, but it's very common to have 2x2 binning (an example from a different one 3840x2160 native, 1920x1080 with 2x2 binning).
https://www.sony-semicon.com/files/62/pdf/p-13_IMX317CQC_Flyer.pdf
To crop the sensor and get back to the 1448 x 938 resolution, they would have to crop it, then stretch the pixels to fit back in that resolution. That makes zero sense to do just to try to hide some HW3 differences, especially given the aspect ratio and the rest doesn't match anyways with the older sensor.
Assuming the AP computer has spare inputs for two additional cameras. And assuming it has the horsepower to use them. But modifying the software would be a bit less than trivial.
It's never as simple as it seems.
Because this is the web, and that's what folks do here.
I think Tesla doesn't care about an upgrade path because they believe that HW3 is capable of driving the car better than the average driver. It sees well enough and has enough processing power, but Tesla hasn't figured out the control system for it yet. They want HW4 in order to expand the operational envelope. To see farther, to be aware of more vehicles, to have better perceptions in inclement conditions, handle atypical scenarios such as hand-directed traffic, and so on.
I figure HW3 will provide Level 3 autonomy, but that there will still be plenty of scenarios that the car will turn over to the driver, such as hand-directed traffic. I just hope that Tesla figures out a good way to alert the driver as far in advance as possible. We don't need blaring alerts to drag us from complete inattention to full attention within a second. Perhaps they could implement some kind of escalating system where we go from inattention to attention to control to attention to inattention. Recognizing hand signals so that the driver could be alerted would be HW3, while the car handling those signals itself would be HW4.
I won't be at all surprised to see Teslas in five or ten years with all sorts of different sensors because they'll be cheap enough at that point for mass adoption. But I do wonder if Tesla will ever decide to put expensive sensors on a vehicle that is specifically intended for commercial robotaxi use. They may be pressured into it by Waymo, Cruise, et al. Alternately, the robotaxi companies may be put out of business by a raft of highly-competitive private robotaxi offerings by Tesla owners.
What about HW3 do you believe is incapable of dealing with hand directed traffic but is otherwise capable of driving better than a human driver?
Do you realize that to "turn the car over to the driver" the car will need to realize it's being directed by hand?
There's a lot of people that don't believe that L3 is really a thing because the very detection of those situations is the majority work in handling them. The real L3 systems will likely be L4 in situations (on the highway) but not in all ODD's, and it will warn you when it's exiting the L4 domain. Not that it's L4 until it isn't.
The processor. There are only so many distinct scenarios that a given set of processing hardware can handle. Computes, memory, bandwidth, etc.
It'll have to realize something, but that something may be as simple as "A pedestrian is standing in the road and won't move". So long as the machine learning system has sufficiently low confidence as to its next move, it can alert the driver. Hopefully, it can nudge the driver when its confidence is just starting to drop. Drivers would then develop a sense of when the car is likely to deal with a given situation successfully.
Who said anything about additional cameras? The proposal was to replace the cluster of three front-facing cameras (wide, normal, narrow) with two front-facing cameras pointed like X-Y microphones, where the one on the right points left and the one on the left points right. It would actually involve fewer camera inputs.
Assuming the new cameras are of similar resolution, the going from three down to two is trivial.
Similar threads
- Question
