Running Computer Vision on RK3566 (Radxa Zero 3W, 3E)

00:00:00 - Intro RK3566 - Radxa Zero 3W
00:00:20 - Cameras for comparison
00:01:12 - How to configure the board for RKNN (system set up )
00:02:51 - Driver patching
00:03:55 - Computer Vision speed
00:06:11 - CPU speed
00:06:27 - Temperature balance
00:06:45 - Wrap up. My opinion.

[verdenum]

Thank you for the video! There is also orange pi 3b which also uses rk3566 and costs simular, but at the same time has much more peripheral interfaces and bigger community support. I think it maybe more friendly for user than the board in the video

[wolpumba]

*RK3566 (Radxa Zero) Board Summary:*

Pros:
- Cheap: Costs $20.
- Standard Linux: Runs normal Linux with existing libraries and software.
- Good NPU: Includes an NPU with a decent inference framework (RKNN Toolkit).
- Good thermal balance: Passive cooling is generally sufficient.

Cons:
- Slow NPU: Twice as slow as the RK3568, despite using the same chip.
- Limited parallelization boost: Only a 20% speed increase through parallelization, compared to much larger gains on other boards.
- System setup challenges:
- Initial setup requires a monitor with mini HDMI or a UART connector.
- USB connection is not reliable out of the box.
- Documentation lacks information about UART connection.
- NPU may require driver patching.

Overall:
- An important step forward in affordable computer vision technology.
- Offers a good balance of price and performance.
- Not the fastest board, but a solid option for budget-conscious projects.

i used gemini 1.5 pro

[zetadoop]

May be it will work faster if you add a heat sink to a chip? It has 5 serials that impressive. Why 1GB of RAM is not enough? I think you was reviewing LUCKFOX with 128mb and 64MB was taken buy IP camera that you failed to disable. (sorry if I miss understood) So my math is one layer aproximately is 512x512x4 = 1MB and next layers can get smaller. So for SSD models it is like 50MB for a model. So why 1GB is not enough?

[ailabinsev]

Maybe low speed is chip binning result

[EvgenSpinin]

Спасибо, с меня лукас!!!! Хотелось бы глянуть в вашем исполнении компьютерное зрение на этом одноплатнике...меня предупреждали, Рокчип, это та ещё головная боль, на днях получил свою радксу, м..да, после распбери не ожидал, что настолько)))

[Zeemer]

Hi, what linux image so you use ? is it Armbian ?

[AlexeyYalaev]

Спасибо за видео! Очень ждём подробный гайд с примером по MAX78000. Невероятно

[alexchukson]

Спасибо за огромную проделанную работу. Какие платы рекомендуете для компьютерного зрения в индустриальном исполнении (-40...+85 С)? Пробовал на Radxa RockPi 3A модель SSD из rknn model zoo. Нужно: сегментирование DeepLabV3plus (backbone resnet50) и из библиотеки timm. RockChip потянет как полагаете?

[greyson]

Hello sir I appreciate your video. I am a big time noob at this stuff but I have one comment and one question.

1: i connected to it and have used it exclusively thru WiFi/ssh, just edit the before.txt in when you make the image. You probabaly know this. I have never controlled it thru hdmi or uart at all personally.

2: the board only supports 2 cameras out of the box, how are you supposed to get other cameras working on here? I have been unable to figure it out. Thank you

[Deciptikon]

Прикольно, а какие еще есть минипк для компьютерного зрения? Пробовали ли вы нейроморфные процессоры?

[unoarecare]

How many fps running yolo from camera?