@CarlK For what it is worth my understanding on this one is the same as Bladerskb’s. My comment is of course not evidence of anything and I have no desire to dig any evidence up but just noting that this is how I recall things as well — Tesla’s camera setup is similar to what MobilEye was pitching and working on back in the day already not the other way around. I agree it would be nice if someone would dig up some materials.
OK we are cool then. I have no idea how Tesla came up with the design either although knowing Tesla's first principle of engineering design it will not do something simply because others are doing it that way. The prime example is Tesla is the only company that did not follow Waymo to use LIDAR to develop autonomous driving. For Mobileye defenders it does not count because all it had been doing was driver assistance.
There's no need to agree to disagree with these people. Mobileye has been pitching its tri-focal camera publicly in 2014 (also privately to auto companies including Tesla) and released samples of its EyeQ4 that powered their proposed 8 camera config early 2015. Moving Closer to Automated Driving, Mobileye Unveils EyeQ4® System-on-Chip with its First Design Win for 2018
static raw pictures, of course. Ah, found it: AP2.0 Cameras: Capabilities and Limitations? and then clarification here: AP2.0 Cameras: Capabilities and Limitations? only ~20 pages ago in this same thread
The dynamic range in those images looks pretty bad. The Sun is bright and everything else is nearly pitch black.
Providing that is directly AT the sun, I say the dynamic range is amazingly impressive!!! I don't think human eyes, or ANY camera that you and I have, can even see the traffic light right at the same direction at the sun. Quoting verygreen's image here: Besides, this picture is just a compressed representation of the picture. Neither the jpg format, or your phone, or our monitor can see that much dynamic range. The different shade of darkness in the lower part of the pictures can be just as high contrast in numbers as in a normal light condition.
Huh, that looks a lot different. I was looking at a another version of the image. This one: Notice how the vehicle directly in front is invisible in this version. The version you posted does look pretty good.
there was a 32bit tiff that you can see full dynamic range in, look at that (use your favorite hdr processing tool!)
Yea, that's the thing. Our monitor just can't show you that much dynamic range. You can use any image processing tool to selectively "look" at a particular range of dynamic range of the picture. Then "we" can "see" better. The detail are all there, just that "we" cannot "see" it. The computer or program can. They are just numbers.
March 20, 2015, Model S wiring diagram was updated with three cameras, rear radars, driver-assist ECU. May 2015, a Model S test vehicle was spotted with triple-camera rig that Mobileye had applied patent protection for (looks like it’s still pending!). In September 2015, news broke that Mobileye was preparing for "first vehicle based on 8 cameras, one radar and ultrasonic around the vehicle." (Electrek, TMC) I made a comprehensive timeline on the Tesla / Mobileye collaboration and break up here
also other companies aswell that amnon pitched to were using the same camera system as you can see here. Autonomous drive technology – trifocal camera Mobileye literally invented the trifocal camera. There's no record of its mention before Mobileye publicly revealed it in july of 2014. Its gonna be interesting if Mobileye starts flexing its patent muscles 3-4 years from now when they start demanding royalties. Tesla literally has no patents so they will be in a huge pickle because they can't offer to cross license. They would have to pay royalties or be sued.
There have been trifocal setups in other contexts before. Fraunhofer HHI presented a trifocal camera system in a different context early in 2014, and Daimler used a trifocal radar in the 2013 Distronic plus. I have a degree of doubt that an arrangement of three cameras (alone) can be patented if it's the only arrangement possible to achieve a defined end. Like it's impossible to parent a two camera setup for achieving stereoscopy, unless additional factors are conditioned into the setup. Same as it's likely hard to patent the general setup of a camera, running its output through a NN and use the result for executing a driving policy - on generally available hardware. With mobileye hardware (EQ chips)? Sure. With general or custom hardware? Not so sure.
This is the Fraunhofer cameras you are referring to which is used create 3D images. https://www.disneydigitalstudio.com/app/uploads/2014/06/MakeBelieve2.jpg Just because its called "trifocal" or uses "3 cameras" doesn't mean its the same thing. The patents are for similar design and functionality. And yes using two cameras to do stereo and depth is patented. I don't think you realize how our patent system works. Everything is patentable. Everything.
Yes you can patent everything in the US. That doesn't necessarily mean a patent can be upheld if it gets challenged.
Here is one of the cameras (7.42 MP) that is compatible with eyeq4 and eyeq5. https://www.sony-semicon.co.jp/products_en/IS/sensor4/images/flyer/IMX324.pdf Sony Releases the Industry's Highest Resolution 7.42 Effective Megapixel Stacked CMOS Image Sensor for Automotive Cameras Versus Tesla's camera ( 1.2 MP) (Datasheet) AR0132AT pdf - Aptina Confidential and Proprietary Preliminary‡ AR0132AT: 1/3-Inc 1/3-Inch CMOS Digital Image Sensor (1-page)
Anybody know why they're cramming more pixels into an automotive camera? It hurts SNR and then gets downsampled anyway.
SNR is only hurt when you use otherwise same level of technology. In a few years camera sensor technology improves significantly (As could be seen in phones and DSLRs)
Higher resolution could arguably help discern more distant detail. Take a photo with one of those 42 megapixel cameras and it is surprising how much you can zoom in.