2 ай бұрын
2 ай бұрын
4 ай бұрын
6 ай бұрын
6 ай бұрын
7 ай бұрын
Пікірлер
@vivekranjan2947 12 сағат бұрын
Hi Anton. I tried running my converted RKNN model on my rk3588 board but am getting an error that the .RKNN model is either formatted wrong or is or corrupt. Can you plz send me your converted model ' yolov8n.rknn' so that I can check my inference.
@AntonMaltsev 11 сағат бұрын
You need convert on corresponding version of RKNPU SDK, rknn-toolkit + export for correct architecture. A random model will not help you.
@Alexnirvan77 18 сағат бұрын
Anton, have you tried to run inference on different NPU cores? (Im having an issue with it, inference works only while using core 0). Also, is there an option to run inference on all 3 cores?
@AntonMaltsev 11 сағат бұрын
On Friendly Elec 3588, Mekotronics 3576, Orange 3588 and Rock 3588 I tested it. It was sucessfully worked.
@littleyogi3341 2 күн бұрын
what challanges we can achive with rk3588??
@AntonMaltsev 2 күн бұрын
all of them!
@r_ruslan 4 күн бұрын
Thank you for sharing you helped me a lot
@andreyl2705 4 күн бұрын
awesome)
@AlexanderDhoore 4 күн бұрын
Anton, living in the future :)
@AntonMaltsev 4 күн бұрын
:D
@q-engineering 5 күн бұрын
Two remarks. 1) In the absence of the GPU, the Rock5C-Lite is targeted for CLI applications. With the official Debian KDE-OS GUI, your CPU load can be up to 6x 85% when showing videos or RSTP streams on the screen. (Not too handy when it comes to vision tasks) 2) The popular Ubuntu OS of Johua-Riek doesn't support the RK3582 (yet).
@AntonMaltsev 5 күн бұрын
Thank you, Rients! Yes, totally agree. CLI + no-streaming.
@aniketpatil4430 6 күн бұрын
where to buy can anyone please send some links to buy theser products
@harishs2651 6 күн бұрын
how can rk3582(4 tops) is faster than rk3588(6 tops)?
@AntonMaltsev 5 күн бұрын
1) It's not 4TOPS, it's 5TOPS according to the documentation - dl.radxa.com/rock5/5c/docs/hw/datasheet/Rockchip%20RK3582%20Datasheet%20V1.1-20230221.pdf 2) It's not only about NPU performance. It's about memory speed, frequency, etc. The result may differ for different nerworks / drivers/pip wheels.
@andreyl2705 7 күн бұрын
awesome)
@flagcrew 7 күн бұрын
КРАСАВА
@ekcrisp1 9 күн бұрын
Great video! At 2:57 what model were you running and how were you running it? Was it the "flusk api" you mention later? Why do you think 76 was faster?
@AntonMaltsev 5 күн бұрын
1) Qwen 1.8 (running with this approach - 08:25 ) 2) github.com/airockchip/rknn-llm/tree/main/examples/rkllm_server_demo 3) 76 may be faster for some models in some conditions. However, more tests are required to verify this. Specifically for the tested model - yes.
@samiuddin1150 10 күн бұрын
We can use Polarize filters and IR Projects with the camera to avoid reflections and glares. Please give your insights.
@samiuddin1150 10 күн бұрын
Like Zed X Cameras with Polarize Lens
@AntonMaltsev 10 күн бұрын
Hi, Sami! Thanks for the feedback! 1) Yes, if you are creating your own camera, it's super important to choose: a) Optical schema polarization, filters, lens coating, etc.) b) Light schema With both of them, you can improve quality a lot. And there are a lot of tricks here. 2) In general, polarized lenses may decrease image quality. For example, if you need to look at a monitor, sky, water, etc., 90% of the time, they increase quality and 10% decrease it.
@ComputerVisionEngineerin-qs1yq 13 күн бұрын
Great Comparison
@mugglepower 14 күн бұрын
my clients would like to run visual recognization on this chip for 1080p images, expecting 2 seconds a picture, would this be realistic? I tried yolov5m on my raspberry pi 5 which is quite a bit faster than the rk35xx and it took it average 12 seconds to do that..
@AntonMaltsev 14 күн бұрын
I think so because RK3588 NPU is much faster than RPi. But of course, it requires a little bit of testing and looking deeper into models/inference pipelines, etc.
@BobRos2 16 күн бұрын
Nice overview about the things which to take into account to choose a right camera. And there were many details I didnt know before. Thanks!
@dwudcuqoyl1465 17 күн бұрын
try Ultra version, and use USB camera, builtin one is garbage has miss-soldred MPIS port, found that the hard way, That said its a €50 ai enabled SBC, with PoE capablities, its not bad for the price range, but would go with the RK3999 Chipset if I got budget
@ПетрВоротынцев-б1ш 17 күн бұрын
Nicely done and gives good overview of camera selection issue. In my experience, camera driver compatibility for RockChip SBCs was the most painful.
@SA-oj3bo 22 күн бұрын
Hi can you help to use yolo11 on rp5 by with the HAILO M.2 ML kit for a webcam and ip-cam? Any codes available that you know? Thx!
@АйдарХузин-м4п 23 күн бұрын
Hi Anton! Thank you for this video. You said that in the end of the day it doesn't matter witch stereocam is used. Could this conclusion be also applied to cameras with structured light? Stereo vision used in realsense d or orbecc gemini series uses active light to project IR pattern. However there is no prior knowledge in the pattern, it just helps to detect feature points. Could you please review or provide your opinion about orbbec astra cameras? I also checked your old video about 3d reconstruction and there realsense d and orbec astra were presented just as cameras using the active light. Maybe there is no significant difference between binocular stereo vision and structured light when speaking about mentioned devices? Thank you for sharing your expertise, much appreciated!
@AntonMaltsev 23 күн бұрын
Привет, Айдар! With modern neural networks, this structured light is not something critical. It can increase cuality in some specific conditions (very big uniform planes, low light conditions, etc.). This camera has structured light. But for 90% conditions, I don't think it's critical.
@boompow2282 27 күн бұрын
U kinda sound like viktor from arcane 😁😁
@romanbolgar Ай бұрын
Ничего непонятно, но очень интересно. Загорелся думал, что можно будет вставить в Nvme М2 SSD И на винде запускать stable diffusion. А тут опять командная строка наверное... И картинки будет рисовать в командной строке... Когда уже доработают эти технологии. Вроде бы востребованы искусственный интеллект и где он? Начинаешь что-то искать кругом только для игр. Видеокарты. А вот такие вот интегрированные модули для нейросетей днём с огнём не сыщешь.
@gergelytoth5613 Ай бұрын
I would appriciate an interface outside of the demo without that watermark and bigger resolution.
@AntonMaltsev Ай бұрын
You can use code from here - github.com/facebookresearch/sam2 Here is my COLAB demo, but I am not sure it still will run - colab.research.google.com/drive/1Un09HITLLM-ljkG1Ehn9cJjdwk8FVI_1
@vswuk-p2v Ай бұрын
Годно
@boltvalley3076 Ай бұрын
Thank you.
@a.b.c4812 Ай бұрын
CRE stereo could be used for metric distance estimation or what could best distance measurement system for self driving car is streo is accurate enough does it provide distance range
@RobertNewport Ай бұрын
Another interesting board worth mentioning is the AMD Zynq 7000 series, which works with the open source Pynq project, which comes with Jupyter Notebook and tons of examples. Though it is not an NPU per se, it does accelerate a lot of ML utility functions and could be a great pre-processor in a low power ML pipeline.
@attashemk8985 Ай бұрын
From my experience, sometimes could be very useful to build prototype in unity and prepare some virtual test environment before building robot
@fluffsquirrel Ай бұрын
Yolo is legendary. Just hope this can run fast on CPU
@tsvigo11_70 Ай бұрын
Наивысшую услугу родине окажет только ЛИБЕРАЛ. А человечесво спасут только специалисты по машинному зрению.
@АндрейТаранов-ь8э Ай бұрын
файнали, ин шорт термс, коллект йоур овн датасетс, трейн зе сота моделс бай юселф, энд хев а фан, экчуалли. Мени сенкс.
@burhanudin8687 Ай бұрын
What is the model of the top section video?
@AntonMaltsev Ай бұрын
It's original output from OrbBec scanner. I expect it should be a little bit modified StereoBM algorithm
@burhanudin8687 21 күн бұрын
@AntonMaltsev did u used monocular webcam or stereo one? I found it interested to try 😄
@pedromunuera7997 Ай бұрын
Please, could you add the links to the models? or papers?
@AntonMaltsev Ай бұрын
For UniMatch I used this version - github.com/fateshelled/unimatch_onnx (original one is here - github.com/autonomousvision/unimatch ) For Depth Anything I used this one - huggingface.co/spaces/depth-anything/Depth-Anything-V2
@samiuddin1150 Ай бұрын
I have experienced a lot of stereo, ToF Cameras and also with Deep Stereo algorithms, I would say that, None of these two can surpass accuracy of Neural Stereo Depth, RAFT Stereo, High Frequency Stereo Matching Network, Stereoformer etc are the whole on another level when comparing with Classical and Mono.
@AntonMaltsev Ай бұрын
Yes, totally agree. I thinkl that for some super specific cases, the plain math can be a little bit more accurate. But for 99% of real tasks, it doesn't matter.
@AntonMaltsev Ай бұрын
kzbin.info/www/bejne/aGXMqKqhqKmZe5Y - Second sample kzbin.info/www/bejne/j3Xdh2ttm8qYraM - First sample
@mal-avcisi9783 Ай бұрын
Why do people always say their names at the beginning of the video, who cares about the name, what does it matter, what value does it add to the video?
@DaengRosanda Ай бұрын
any video updates regarding of this?
@Orbbec3d Ай бұрын
Thank you so much for testing our camera and making this awesome video! Your review means a lot to us and really helps others see what the product can do in real-life scenarios. We truly appreciate your support! ！！
@Orbbec3d Ай бұрын
"Thank you again for the video! I just wanted to mention that the product name in the video title might have a small typo - our product name is actually 'Orbbec Gemini 335L.' Just a friendly note in case you'd like to update it!".
@AntonMaltsev Ай бұрын
Thank you!
@mercury3444 Ай бұрын
Где тогда получить этот опыт начальный для вката в машинной обучение?
@mercury3444 Ай бұрын
Кто-то говорит что ии всех скоро, ии же очень выгоден для бизнеса но вакансий в сфере машинного обучения очень мало. Почему так? Должен быть скачок
@tylerdavis5796 Ай бұрын
Fantastic info! Will probably look towards another solution for edge inference now
@AntonMaltsev Ай бұрын
Here is the link to the article on how to run everything - medium.com/@zlodeibaal/how-to-run-hailo-on-arm-boards-d2ad599311fa
@tomlongson5903 Ай бұрын
What would you use for the desk scene for pick and place?
@AntonMaltsev Ай бұрын
Nice question! It depends on a lot of questions: What is the gripper? If it's some vacuum suction cup, all cameras are OK for objects with plain surfaces. If it's a human-like arm, I may train my network on raw stereo data. For example we did it like this here - kzbin.info/www/bejne/gKSZkJh3a8hkhNk What is my budget? With an unlimited budget - I would prefer Mech Mind or some Zivid cameras. It's much easier for a complex objects - kzbin.info/www/bejne/oaeThml8bNB4nsk Can you install a 3D camera on your hand? D405 can be nice at a close distance. Also, I would think if I have enough computational power to run stereo reconstruction myself - kzbin.info/www/bejne/o3bWkGusrd2Gp6s Sometimes it can give the best quality
@AlexanderDhoore Ай бұрын
I'm a researcher from Belgium at university. Your channel is a true gold mine. Keep up the brilliant work.
@ПетрВоротынцев-б1ш Ай бұрын
Thanks for the great experiment, I was just planning on buying the new Pi-AI-Hat but I'll have to change my mind.
@AntonMaltsev Ай бұрын
It will probably boost your speed a bit, but not for all models and probably not as much as you expect. I saw some reviews that usual boost is 30-40%, and specifically some slow models it could be almost x2.
@ARMENIA181 Ай бұрын
Thanks for the nice video. Is it possible to connect two cameras to the RK3588 and have depth estimation? I mean, does it require image synchronization?
@AntonMaltsev Ай бұрын
Hi! 1) It is better to have a stereo camera with onboard synchronization. Without it, you can't work with moving objects. If they are static, then it should be OK. 2) It's hard to run some workable network on RK3588. It was pretty big part of this video:)
@BobRos2 Ай бұрын
Nice video, as always! I would be happy to see also something about mono cameras here, which would be used in local robot setups, for example with robot arms using ROS or used in other areas like small rooms. There eare so much camaras ou there that it's hard to find good ones.
@AntonMaltsev Ай бұрын
Thank you! It's a nice idea, but it's pretty hard to do. There are a lot of them, and it's super hard to do a good comparison. I have a plan, but definitely not in the next few months.
@andrew6274 Ай бұрын
друг, почитай про пленоптические камеры и вообще про камеры светового потока И анубисом тебя заклинаю, поправь произношение
@blackbigdeath 2 ай бұрын
Норм, только надо шрифт покрупнее чтобы был, и большая часть по времени таких видео должна быть занята схемами, примерами, это всё-таки для технарей. Норм ❤
@saitaro 2 ай бұрын
Отличный ролик и канал, спасибо, Антон!