Raspberry Pi Launches Camera Module For Vision-Based AI Applications (techcrunch.com) 15
An anonymous reader quotes a report from TechCrunch: Raspberry Pi, the company that sells tiny, cheap, single-board computers, is releasing an add-on that is going to open up several use cases -- and yes, because it's 2024, there's an AI angle. Called the Raspberry Pi AI Camera, this image sensor comes with on-board AI processing and is going to cost $70. In more technical terms, the AI Camera is based on a Sony image sensor (the IMX500) paired with a RP2040, Raspberry Pi's own microcontroller chip with on-chip SRAM. Like the rest of the line-up, the RP2040 follows Raspberry Pi's overall philosophy -- it is inexpensive yet efficient. In other words, AI startups aren't going to replace their Nvidia GPUs with RP2040 chips for inference. But when you pair it with an image sensor, you get an extension module that can capture images and process those images through common neural network models. As an added benefit, on-board processing on the camera module means that the host Raspberry Pi isn't affected by visual data processing. The Raspberry Pi remains free to perform other operations -- you don't need to add a separate accelerator. The new module is compatible with all Raspberry Pi computers.
This isn't Raspberry Pi's first camera module. The company still sells the Raspberry Pi Camera Module 3, a simple 12-megapixel image sensor from Sony (IMX708) mounted on a small add-on board that you can pair with a Raspberry Pi with a ribbon cable. As Raspberry Pi promises to keep production running for many years, the Camera Module 3 will remain available for around $25. The AI Camera is the same size as the Camera Module 3 (25mm x 24mm) but slightly thicker due to the structure of the optical sensor. It comes pre-loaded with the MobileNet-SSD model, an object detection model that can run in realtime.
This isn't Raspberry Pi's first camera module. The company still sells the Raspberry Pi Camera Module 3, a simple 12-megapixel image sensor from Sony (IMX708) mounted on a small add-on board that you can pair with a Raspberry Pi with a ribbon cable. As Raspberry Pi promises to keep production running for many years, the Camera Module 3 will remain available for around $25. The AI Camera is the same size as the Camera Module 3 (25mm x 24mm) but slightly thicker due to the structure of the optical sensor. It comes pre-loaded with the MobileNet-SSD model, an object detection model that can run in realtime.
Bad tech writing (Score:5, Informative)
"AI startups aren't going to replace their Nvidia GPUs with RP2040 chips" is extremely misleading. The IMX500 does the AI work. The RP2040 is just an interface adapter.
Re: (Score:2)
Re: (Score:3)
Yes. And the IMX500 only seems to have 8 MB of memory for everything (firmware, model weights, etc.) so any old GPU should work fine for training any model you could fit on the thing.
They are becoming self aware (Score:2)
Onboard processing for 3D printing? (Score:3)
This is definitely useful for onboard Obico checking of 3D prints in progress to check for layer shifts, prints not sticking to the bed, pasta, or the start of blobs of death.
Yes, Bambu printers do this, but it is nice to have a F/OSS version, preferably with the LLM hosted locally for security reasons to go through and ensure things don't break in a crazy manner.
Sweet sensor, but unknown AI (Score:5, Informative)
The sensor is pretty sweet, 12 megapixel supporting 4k60 and 2k240 video modes. Weirdly there are no specs on the AI processor. I'd view this more as a way to get the IMX500 sensor if you're interested in that. If you're interested in embedded AI, you might look at the Luxonis options, which work with the Pi over USB.
Re: Sweet sensor, but unknown AI (Score:1)
Itâ(TM)s a DSP with 8M of SRAM. It can do simple things like segmentation on relatively low resolution (across the sensor you have to work with VGA or less resolution), and you canâ(TM)t stream video at the same time. So in certain crude well defined applications such as anomaly detection on a factory conveyor belt, this may be able to work.
Re: (Score:2)
Some here:
https://developer.sony.com/imx... [sony.com]
Well, there goes ... (Score:3)
Why 1 TOPS when already RPI 13 TOPS speed? (Score:2)
The SONY IMX500 is about 1 TOPS. The existing RPi kit with Hailo-8L NPU found in the Raspberry Pi AI Kit is 13 TOPS.
https://www.raspberrypi.com/pr... [raspberrypi.com]
Price? $70 versus $120 plus RPI5 cost?
Power consumption?
This is a competitor to the Sipeed MaixCam: https://wiki.sipeed.com/hardwa... [sipeed.com]
Sipeeed have been making this stuff for a while... nice little modules, but the English documentation is a little bit thin sometimes....
Re: (Score:2)
Or the Jetson Nano, at 40 TOPS:
https://developer.nvidia.com/e... [nvidia.com]
Re:Why 1 TOPS when already RPI 13 TOPS speed? (Score:5, Informative)
1. TOPS are like MIPS. You need to measure actual performance, not just multiple number of MACs by the clock rate. You network has other parts beyond the MAC units, and if you are not efficient, it is going to slow you down, no matter how many TOPS/s your MACs do, similar to Amdahl's law for multiprocessing.
2. The IMX500 has been announced on May 2020, and has a mature toolset. It's just this specific RPi product that is new.
3. Check the TOPS/W, not only TOPS, but even TOPS/W numbers are skewed (https://semiengineering.com/lies-damn-lies-and-tops-watt/).
4. Even better, check actual benchmarks: https://arxiv.org/pdf/2307.078... [arxiv.org]
Re: (Score:1)
Yes, good points. I had not thought this through yet. I will get a Raspberry Pi AI camera when they are generally available.
One weak point of the IMX500: it only has 8 Mb of memory, limiting what models you can run.
https://developer.aitrios.sony... [sony-semicon.com]
Sipeed had modules like this two or three years ago. They have moved on to make more useful cameras like the MaixII and MaixCam. The MaixCam has 256 Mb ram and costs ~$70 with screen and camera, and ~$40 with only camera. The 500 mA (max) power is not too bad, si
I could use this (Score:1)
I don't work there anymore though, and they seem largely uninterested. An rPi with