How good is OAK-D Lite as a 3D camera?
ฝัง
- เผยแพร่เมื่อ 2 ก.ค. 2024
- 0:00 - About OAK-D Lite
1:04 - Unboxing
1:49 - Termal
2:58 - Depth Quality
5:20 - Comparing depth with RealSence and Zed
5:50 - First Way to increase Depth quality for Oak-D lite
6:28 - Second way to improve Depth quality for Oak-D lite
7:51 - Third(?) way
8:16 - Closing words
And, do not forget that we provide CV-related consulting - cvml.rembrain.ai/
Subscribe on my LinkedIn - / anton-maltsev-037a5684
Semi-global matching algorithm - en.wikipedia.org/wiki/Semi-gl...
Smoothing algorithm - github.com/luxonis/depthai-ex...
Interesting issue - github.com/luxonis/depthai/is... - วิทยาศาสตร์และเทคโนโลยี
Yes, thank you for the review. Quite fair and well explained I think. (Brandon from Luxonis here.).
We're working on some internal efforts on using neural inference to improve the SGBM results. Which as you allude has the promise to make this an insanely good depth camera.
We're also more than happy to send you an OAK-D-Pro to try out. As this has active stereo as well which helps with blank surfaces/etc.
Thanks again,
Brandon
Thank you Brandon!
You're doing an amazing job, and your devices are definitely redefining what Computer Vision should look like.
I would love to test your new camera. But in the next few months there will be problems with sending it to me. Can you email me? I will describe the options how it will be possible.
My e-mail is maltsevanton@ya.ru
@@alfred8294
1) Brandon did not pay me anything for this video. This camera I brought myself for my own project. And I am happy with it. Yes, they sent me the next camera for the test for free. But at that point, I brought 2 cameras myself. And for now, we use around 10 cameras in different projects.
2) I know about 5 startups happy with the OAK series.
3) Brandon has passed away this year.
OAK is one of the best user-experience cameras. You will have many problems with other cameras if you can't handle OAK.
@@alfred8294 I used it on:
1) Windows (3 different PC)
2) Ubuntu (2 different PC + Jetson (a few of them))
3) RPi OS
Also, my friends using it on Android
The last time I used it was ~3 months ago. On Windows PC (AMD CPU, NVIDIA GPU)
Very good video!
Very interesting review. You answered my main question - approximately what depth resolution should I expect from my robot’s Oak-D-Lite in the 0-1 meter range for input into my home mobile robot SLAM calculations. For my autonomous mobile robot the Oak-D-Lite offers a very appealing 3.5W power use, at a very nice price, and small size/weight - hopefully it will be “good enough”.
How is the vibration handling in autofocus oak-d-lite when it is set to fixed focus in software?
Thanks for this, it would be really interesting to see your results on the stereo depth neural network. I’m curious, why do you mention that monocular depth estimation is useless? Is it not able to spit out accurate absolute depth?
Yeah. It's very dependable on image. For example, for this video - habrastorage.org/getpro/habr/upload_files/0a9/ac5/426/0a9ac54267abdfdde4affdf9c86b865e.gif
The result will be pretty fun - habrastorage.org/getpro/habr/upload_files/6a7/254/88b/6a725488b6797c5cff66edabdceab4ea.gif
I used MiDaS for this (rank 2 on paperswithcode ).
Such problems give unpredictable results for real tasks.
I see, still lots of work to do on mono depth estimation…
can you really get 200fps indepth? i want to be able to track a ping pong ball with this camera, should i buy it? realsense has much lower fps! Wonderfull review my driend, thanks alot
Мовидиус же там к чему то еще подключен? Он же не есть основной процессор?
Да, там логика достаточно забавная.
В коробке, по сути, три камеры + movidius. Никакого процессора нет.
Но, софт дает возможность сконфигурировать поток данных таким образом, чтобы данные с камер шли напрямую в movidius. Чтобы host-устройство, куда эта камера втыкается, получало только набор детекций/карту глубины/классификацию изображений.
При этом, если детекция многоэтапная (например детекция -кроп-классификация), то данные надо катать на хост-девайс и назад. Например: камера->miriadX(детекция)->host(кроп)->MiriadX(классификация)->host
Тобишь как обычный movidius его можно использовать. При этом, по простым задачам, где нет промежуточной логики, он может работать даже на очень слабых host-машинах, слабее чем первый RPi.
@@AntonMaltsev Так а карту глубин хост что ли считает?
@@ra9mgk нет. Базовую из коробки Movidius.
Но, если её качества не хватает:
5:50 - First Way to increase Depth quality for Oak-D lite - вот тут показано что будет если глубину сгладить на хосте
6:28 - Second way to increase Depth quality for Oak-D lite - вот тут показано как скорее всего через несколько месяцев можно будет считать глубину на OAK-D
@@AntonMaltsev А обычный OAK-D ведь тоже дает возможность прямого потока данных с камер в movidius? Просто не совсем понял, почему Oak-D Lite интереснее позиционирует себя чем обычный Oak-D, в последнем ведь железо получше будет?
@@andreyskourikhin5768 OAK-D Lite в 3 раза дешевле, в 3 раза меньше.
Разница там только в IMU модуле как я понимаю.