Sharper than the iPhone!
ฝัง
- เผยแพร่เมื่อ 18 พ.ค. 2024
- 64 megapixels and autofocus for your Raspberry Pi. Can you really resolve that much detail in such a tiny camera, or is it just marketing?
Support me on Patreon: / geerlingguy
Sponsor me on GitHub: github.com/sponsors/geerlingguy
Merch: redshirtjeff.com
2nd Channel: / geerlingengineering
#RaspberryPi #ArduCam #Photography
Mentioned in this video:
- ArduCam 64MP Hawk-Eye camera: www.arducam.com/64mp-ultra-hi...
- My blog post with setup details: www.jeffgeerling.com/blog/202...
Contents:
00:00 - 64 megapixels!
01:03 - Camera specs
03:40 - Image quality and autofocus
06:45 - Video, continuous autofocus, and digital zoom
08:11 - Jeff's take - วิทยาศาสตร์และเทคโนโลยี
![Supernova - aespa エスパ 에스파 [Music Bank] | KBS WORLD TV 240517](http://i.ytimg.com/vi/bepqfR-pxY0/mqdefault.jpg)
64MP ought to be enough for anybody!
Where have I heard that before? Didn't IBM say there was only a need for one super computer? 64MP is enough until 128MP is on the market.
Nice try.
it's the sensor size that matters
@@tonysheerness2427 108mp was on phones in like 2020 lol
@@tonysheerness2427 BIll Gates with 640kb of ram is what you're referring to
Love the great info using your knowledge of photography. I always struggle to explain to folks why more MP isn't always better when you're talking about small sensors.
I don't like that every modern smartphone has more MP and more cameras (2 is ok but i don't need a main camera, a wide angle, a macro, and a night vision). I'm ok with 1 camera, not more than 16MP (cheaper, and i don't need more)
The simplest explanation is that larger pixels are less prone to be disturbed by electrical noise and interference within the image sensor. That's one fundamental reason why large sensors are much better in low light level conditions, the other being that they can quite literally capture more light.
Hey Eric, many RPI-based/embedded applications can actually benefit from a small sensor with extra MPs.
@@ArducamOfficial Be great to see some examples of this
These days there's a second trick though... by boosting the pixel count you gain more information to send into image processing AI/algorithms. So even if there are extra artifacts they might be easier to correct versus trying to upscale.
We need manual focus so it can be controlled with some extra hardware like an encoder.
Hey Sean, the focus motor of this module is manually adjustable through software!
@@ArducamOfficial What's the best way to fix the diffraction haze issue?
need an adapter from Pi to full-size camera lenses
@@QualityDoggo the crop factor will be massive
@@QualityDoggo Why? Its 1/1.7 sensor... you need much higher resolution lenses for those tiny pixel. Well, in reality no lens on earth can resolve them.. but even in binned pixel mode, you need lenses made for this sensor size.
i think it would be pretty interesting to build a small photo-scanning-system with this.
to bad i cant program.
You might try skanlite?
scanning? with what? for scanning old analoge films you are better off with full frame camera internals modified as scanner, and if you talk about final images in 10x15 or bigger... the results will be also "bad" and you are better off with a 50mm-100mm prime lens on any camera.
i bought longer ago a broken professional DSLR and torn it apart to use sensor and processing hardware, in addition with a sled, gearings and little bit tinkering i made partially automated for up to 35mm film digitalizing machine, as the super expensive models which are used professionally to convert old movies to digital (just way less automated and probably worse ofc)
The end result were a ton of raws, inverted the colors and just made a movie out of it, each image a frame.
I can't program either. Would you be interested in partnering up with me to abondon the project before we even start it?
Learn how to program
I'm glad you acknowledged the limitations inherent in tiny (and cheap) lens and sensor systems.
Yea but he ignored how many people bought this and never could get it to work. Arducam left a ton of people hanging. In the end, it’s a buggy pos.
I am once again asking people to stop referring to quad-bayer sensors by their sub-pixel resolution instead of filter resolution.
Omg the world will end because of improper vocabulary 😅😂
Ikr
I think one advantage with this is that for 1080p video you can still punch in with the digital zoom to get a tighter framing without really sacrificing resolution. You should still have 1:1 pixels.
binning would still yield better results (unless you're stacking) due to how noisy sensors are at this pixel size.
Man!!! This was a hell of a benchmark! Congratulations! Videos like this are something that makes not regret to subscribe your channel. Continue the good work, Jeff. We keep watching, even if you don't answer messages or comments.
That's really amaizing camera! Great video
I’d been eyeing this guy for a while and I’m grateful that you were able to show its performance.
I was already looking for a pi cam for my octoprint set up, so I pre-ordered it right away! Glad to hear more info on it thanks!
I was watching this in my university library and it turned out that my headphones were not connected properly so the entire library just hear Jeff Geerling scream "64 Megapixels". Thanks Jeff.
HAHA nice. Sorry library.
I'd love to see you make an inexpensive pi based macro microscope for things like soldering
Nice content from you as always! Thanks!
The CM4 has 4 lane CSI-2 hence you can transmit 4k30 video over it.
And the PCIe bus can also barly keep up with 4k30 to allow you to process the video on an FPGA instead of software.
Good analysis, and sound reasoning
Thanks, Jeff. I think I'll use a couple of these for the seeing eyes on my Lathe and Milling machines. Good thing I have a stash of Pi's and Compute modules before the dark times kicked in. Who knew!
Great video. A Few camera manufacturers started pushing micro 4/3's sensors with outrageous claims that they could outperform full-frame Nikon and Sony cameras. A few of my friends fell for it before I could reason with them. They could not understand how my Nikon D810 from 2014 could produce a better image, especially in low light than an Olympus from 2022, and cost about the same as my D810 in 2014. The Olympus does have a better pixel pitch (3.36) vs D810 (4.88) but the D810's sensor is 2 times the size .
Just never believe marketing or "reviews" of populistic "youtubers" (not photographers!)
A good example... a friend once wanted a good (!!!!) surround system... i have much experience with so we tried to find something. We found.. a nice Klipsch 5.1 system with big floor standing fronts, a big center and shelf speakers in the fitting driver size as the fronts and center,... 2nd hand "just" 1800€
Some days later... he said 1800 was too expensive for "old" (built 2008-2011 around!) speakers so he bought a new premium soundbar for over 1000€...
Well... he tried to whitewash his stupid decision for years. Since a while, he admit it sounds bad and is not even close to worth 1000€+.
I mean... thats what happens if you trust a channel which is just reviewing china crap and smartphones. Should be a sign already if someone "testing"/reviewing a soundbar in a completely UNDAMPENED room and is showing the sound off with faked 3D sound so headphone listeners can "hear" the surround...
The sad thing.... "non audiophile" youtubers even reviewed like this 3000$ Soundbars like the Sennheiser Ambeo... As if i would test my huge 7.2 setup in a 10qm room barely fitting the speakers with me at all....
It's a shame the bandwidth of the Pi limits the video ability. The image resolution isn't that important really. Shame it also doesn't have pixel binning, that would help reduce noise at the expense of resolution, which I think would be worthwhile.
Also, the Hawkeye seems to render skin tones differently - seems to have a magenta tint. I'd say the normal Pi camera has nicer skin tones. Not that it's probably meant for portraits anyway.
Second on that pixel binning
Yeah my 27mpx shots on the 108 on my phone is more than enough and makes up for the small pixels
Hey Izzie, it does have pixel binning (2×2)!
you could get an Orange Pi.
Oh my this will be amazing for Octolapse timelapses
Great review!
Great Project, great Observation and great explanation. 😀👍
ok but that squier in the background looks so good
You are an awesome man, i liked your work. Thanks for the video.
I appreciate your honest caveat about why I would want to do this and your comparison to an older Nikon D700. So many people throw around megapixel numbers as though that is the end all and be all of photos. It isn't. Thanks.
Love the t-shirt ! Oh and the video 😁
Awesome presentation. The explanation of the different factors that make up a good image is usually missing as people swoon over the number of megapixels.
Love the video so far (:
pretty cool! the iPhone cameras make tiny movements also, sometimes when they break they'll get stuck shaking back and forth and it can actually get quite loud!
I’d like something like this for my bird feeders, the problem then becomes power as my ESP32 camera runs off batteries and the flat's (apartment to my American friends) WiFi. As a really enamoured Pi Fan Boy I think there’s still situations where micro controllers have an upper hand.
This looks great! I think I'll try one of these along with some wifi-based video streaming and a battery, so I can set up a cheap multi-cam live stream for the church.
EARLY Keep up the good work Jeff Geerling
Good review, lke your work
One thing to keep in mind that this is a quad-bayer sensor. In theory you only have around 4 MP color resolution, anything above that depends on how good the interpolation algorithm can correlate the different colors. Depending on the captured image, it is no better than a traditional 16 MP camera which also has the same 4 MP color resolution (because trying to make it sound better than it is is nothing new).
The spec page says it can output as RAW10, does this also work for 64 MP mode, or does it revert to 16 MP? My phone with a 64 MP sensor will only give 16 MP bayer data if I attempt to shoot RAW, and I'm really curious how the actual quad-bayer data looks.
The banding seems likely due to the CMOS sensor's readout train. The perfect regularity of it suggests the ADCs to me. Try running the Pi off of a battery or a UPS and see if it still gives that banding.
I've seen similar banding with CMOS sensors used for astrophotography and it's frequently resolved by feeding it clean power. Though sometimes it's a temperature issue (astrophotography being an extremely low light endeavor, thermals contribute significantly to electronic noise in the image).
EDIT: It sounds like Arducam confirmed it's the denoise filter at very high resolutions. I guess the easy way to find out is to disable the filter above 16MP and see the result -- I don't have time to test it out myself for a few days, though.
It's the denoise feature being used at resolutions above 16MP causing a glitch.
Ive noticed that too. Running a Sony A6400 gives much noisier images when its 10 degree winter night vs when its 80+ degree summer day.
To me, it seems like the artifact of 60Hz lights in North America, as normal lighting flickers at 60hz, the average human cannot detect it (although iirc lights now don't do this as much), this flicker can cause issues with the camera
could it be the flickering of 6o/120 Hz lighting from AC powered fixtures interacting with the scanning frequency of the sensor?
Love the title and the explanation on the lens and those comparisons!
When are you gonna share your 4K upgrade? I’ma use this video to convince our marketing team to upgrade their equips…
About those problems:
Banding - Caused by the denoise function (handled by RPI’s ISP) with resolutions higher than 16mp, we have reported this to the foundation, for now you can just turn it off with --denoise cdn_off . (won’t affect image quality)
AF and Continuous AF working weirdly:
The preview version of AF we developed segments the scene into 15 (3x4) areas and uses only 2 in the center to calculate the focus_value and to trigger AF whenever that value changes, this is also the version we sent for review.
The second version, which we’ve been working on, will segment the scene into 192 (16x12) areas and use the entire scene for the focus_value calculation along with movement detection to decide when another AF should be triggered.
just clicked on this video, and my first immediate thought with this was it would be so, so cool if someone made a kit for raspberry pis that turn it into a cute lil DIY digital camera; or hell just some software that combined with a touch-screen, could turn it into that
it's something i'd buy in a heartbeat lol
Very good usecase: Focus stacking. The focus motor can be "manually" controlled. It would be easy to write a script that takes pictures at different focus settings and then stacks them together. Disclaimer: I haven't tried it yet. I ordered one of these, but I hope it works.
i have done that with the 16mp version, it works well but the pi4 does not have the power to run a fcus stacking software at worthwhile speeds.
Hey Jeff, thanks for another awesome video!
Something I would be really interested in watching (if possible with todays hardware), would be if you could thinker with a setup where you would have a camera recording in 4K30 or 4K60 through something like a capture card connected to the CM4.
It seems like the Rpi/CM4 can only manage 1080p 30 fps
@@Jobey_99 you are 100% correct in regards to the normal camera interface, so you would need something external that can capture the video at a higher resolution and then encode it for the CM4, hence something like a capture card :)
Love the T-shirt!
I needed to comment so badly that I haven't even watched the video yet!
Hope you & your family are well
Would be nice if you could digi zoom in sensor, so instead of having a 1080p30 full frame image you could choose a portion of the image at full resolution, cropped and zoomed in to 1080p.
That would make the digital PTZ much nicer.
You can run a 50ma laser over the sensor to rebake the gutter so off center hits drain into the gutter, just go back and forth until only the traces are reflectors. Way back in 1982 they invented the terazertz 186, that should be enough for a music continuous run portrait.
It's mind blowing how far this tech has come...
That seems like a amazing camera for DIY projects, im really interested in building now a automated sentry turret for airsoft or paintball.
ok, this is the best camera for the raspberry pi 4 b what i have ever seen.
Would have been perfect for a machine vision project I am working on if the sensor had global shutter.
Supposedly ArduCam is working on a camera module with a global shutter. Not sure when it will be a reality though.
@@JeffGeerling The Arducam OV9281 has global shutter. It's B/W though but otherwise nice image. There's been some issues with that one as there was only Arducam closed source drivers but I think now maybe there's kernel and libcamera support? EDIT: there should be OV9281 modules from other manufacturers too.
Hey Tokiomi, we do have a couple of mono/color global shutter cameras.
Super interested in these Pi cams.. They seem like a perfect fit for a model train FPV filming kit. I was amazed at the high price and awful quality by most of the spy cam and drone cam setups that are common.. It would be so much more extensible and compatible with a Pi and some recording/streaming tools (AND CHEAPER! $100+ already for Lionel's CupolaCam, and horrible quality compared.)
4:23 There is also some extra color information in the 64px.
If you look near the top center where there are those contacts stepping down the 64px shows shading and color where the 16 doesnt. But even here there is banding it appears like there are black lines on the circuit board.
Hey Jay, it's a bug caused by the denoise function in RPI's ISP, you can fix it by simply turning denoise off.
Unless someone has already mentioned it, that yellow banding is definitely from the lighting which is affected by AC electricity/shutter speed sync issues. I've seen it before in my own pics, with various cameras, using different light sources. It's also clearly identifiable in copy, pasted & zoomed up segments of images and when checked with a histogram. Dem numberz don't lie!! 😉😂
Another channel that reviewed this found the banding in shots that were taken outside in natural lighting too, so it might be something else entirely. Hopefully it can be addressed through a driver patch
th-cam.com/video/TV7DLhQYSmI/w-d-xo.html
@@Jaaxfo well shutter speed sync issues aren't dependent on light, but can affect the light introduced to the system. But looking at the link you provided, there's an update on what's causing it. The denoise feature is glitching when taking images at above 16 MP, so until it's fixed you get around this by turning the denoise feature off for anything above 16 MP
@@epmcgee Thanks for that insight
Yeah, it's almost never limited to artificial light sources when this happens. But yes, it's usually the result of either the shutter speed or any filter (hardware or software) refusing to play nice with available light frequencies.
Interestingly enough, this also tends to be uniquely an artifact of CMOS and, less often, CCD sensor cameras. You *could* capture banding in images on film, but you had to do it deliberately through a combination of special equipment and special techniques. It's one of the reasons why there are still purists who insist that film is the only proper media for photography.
Good info and good shirt lol
Wow, just bought it last week with Arducam's servos pan&tilt kit, for a surveillance cam...Nice timing, that video will help me setup that new cam on a pi4 ! Tx :-) I know it's more editing time, but i miss the bloopers! ;-)
What would be really neat is if the ePTZ could be coupled with the full resolution from the camera so you don’t lose any detail when zooming in X amount
My a7siii is only 12mp and I have a 24mp a7iii and sometimes depending on the situation, I like the siii better. And that’s because it has a faster sensor readout and has larger pixels so in some situations, it looks better for very specific things. Not to mention the AF has better tracking despite them being from the same relative generation. The s model was 2 years newer.
However larger pixels doesn’t always translate into better lowlight. Sometimes having higher resolution means the noise in the image is finer and therefore easier for programs to remove. Larger noise chunks can mean that when an image is smoothened out in post, it looks more blurry.
It’s all very situational even on the high end when you’re spending 2-3.5k on a body. Not including lenses.
I’ll be sure to try it out in Late 2023 when Raspberry Pis come back in stock
I noticed the missing guitar string in both pictures, but the color and detail were clearly better on the Hawk-Eye; as the V2 image looks to be more saturated when you look at the neck, and more blurry when you look at the writing on the head.
If it fits for low light situations surveillance and tracking might be a good task to that
Oh man I really like the idea of building a raspberry pi 360 degree camera. It could work probably well enough with several low resolution camera modules. Just merging the image & getting the optics right so you can capture a wide enough angle sounds quite challenging :/
Another interesting video.
One point about the issues to run the camera at higher resolutions (which I guess you might know already): I had similar issues with the 16MP on my Raspis, since I wanted to run it at full resolution and take still pics at 2 fps at least. I had to work on the cma settings, increase the cma size in the config file to get rid of those memory errors. Did you try this already on the 64MP version?
I still have this dream of making a turntable 3D scanner where the turntable would spin, and a camera would start at the apex and gradually descend with each rotation. If I can get my head around it further, A 4-camera setup with this thing would be pretty cool.
That shirt is the best!
How does it perform in low light compared to the Starvis? And what is the power draw?
Hi Jeff, I notice a judder effect during the panning shots in the intro (0:26 for example). Do you by chance shoot at a different frame rate than your final video output frame rate? If it is so, this could be causing Frame Rate Conversion to kick in which creates this judder effect.
Good to see arducam coming up with these cameras....Hoping to see a camera with OIS soon.
Btw which amp do you use with your Stratocaster? Would love to see a video of you playing it...
Octolapse without setting up my slr? I would like that very much. Yes I know the quality wouldn't be great but this is nice and small. Thanks for the info and keep up the great work!
Ordered already some time ago. To create timelapse videos & monitor 3D prints
Awesome
Good job🤨👍
That is an interesting raid setup you have there,
4 raid cards, 4x4x4 sata/s ports, and 4 molex power outputs on the board it's plugged into, and I guess USB 3 connector is for PCIe signal?
@jeff not sure if this is mentioned before, but by using libcamera you are using memory from the CMA where as the default 32MB might not be enough for >1080P
Can you increase the CMA allocated memory and give higher resolutions another go?
Great video! Thanks for sharing the knowledge.
Quick question, I have a lab with 20 3D printers running all the time and need to keep an eye on any failed prints. How would you recommend structuring the setup if I buy 20 of these cameras? How many of these cameras can be connected to a single r-pi to stream or capture photos every 5 seconds individually? Do I need 20 P4s? Cost is a big factor here. :)
Finally someone is honest to upgrade an iPhone to its full potential :)
This would be awesome to have on your main pc as a webcam. Imagine what you could do with a pc instead of a pi. I WANT TO SEE ITS LIMITS.
well I mean a pi is just a Linux pc so you could just install the drivers on your pc
assuming that there was an x86 driver or the drivers were open source and could be compiled for x86
not really, its a CSI interface camera, so not usb or something normal
"good" sensors have something called global shutter, where the sampling of pixels is sort of frozen in time, where normal cameras rely on rolling shutters. A method like scan-line where each row is sequently read. This looks a bit like an interlaced read-out vs. the stroboscopic effect of the lights vs. the camera read-out.
If you're talking about the banding it's down to a glitch in the denoise feature at resolutions above 16MP
I got 2 at the early-bird price at the early bird price and they just arrived. I'm excited to start playing with them. One will almost certainly be my new 3D printer's cam.
Good luck!
@@JeffGeerling I wrote that (poorly) in a rush. To clarify, the printer isn't new (it and I have been working together for 2 1/2 years now), but one of these Hawkeyes will be the new camera for it.
I like that the cameras are high enough resolution that it can digitally zoom and still give a high resolution image/video. I'll have a clear image whether I'm making a big print or a small one.
@@davidg5898 Yeah, should definitely be enough resolution-just make sure you get it focused decently, otherwise you'll zoom in on a slightly blurry print! I'm assuming you're using Octolapse?
@@JeffGeerling Octolapse, yes. I like to give the video also when I make a print as a gift.
But I also like being able to check on print progress without having to physically be by the printer.
I've been using a USB webcam that I haven't been all that happy with.
The Broadcom GPU on the Pi does 1080p@30 at max according to its data sheet.. You will have to get something like a NVidia Jetson Nano to get 4k -- it also has a 15-pin-connector for camera but IDK about the drivers.
This video was great!
I was wondering how this compares to the HQ PI Camera?
I want to build a remote camera rig to take pictures of a 3’x4’ surface from about 6’ in the air.
Jeff you should try to utilize the hawk-eye to it's full potential with the new pi 5
That's a whole stack!
Minecraft for those who don't know
Are Pi cameras like Hawkeye connectable to a standard computer to use it’s stronger processing power?
4 cams at 1080 (or better) all being record Ed in sync ?
Great for monitoring my 3D printer
Epic.
Would really love to see firm dates on delivery. I ordered a two a couple weeks ago, but the lack of a firm date makes me a little uncomfortable. Also the constantly-resetting pre-order window is a little weird, always at least a couple days left.
cant wait to build a 3d scanning rig when there's finally more pi boards in stock for a reasonable price
Are there similar camera modules with fixed focus? I'm looking at prototyping a streetview type camera with RTK GPS that could be put on fleet vehicles as a lower cost alternative to sending out vehicles specifically for capturing street level photos, and also capture some unconventional image types (simultaneous stereo) to make the output more useful for photogrammetry and SLAM.
Its is not only about the hardware its also about the software thats why smartphone are taking equal or sometime better picture than dslr
Would be interesting to see if the camera can do a 1/4th of the 64MPix image zoom, like, say you have it set up as a security camera at your door, and see in the top left corner someone fiddling with your car. At 1080P you can't make out much details, but setting the camera to operate like a 1080p video, streaming only the first quadrant of the camera, you might be able to see a lot more, without limiting the pi bandwidth. The pi just thinks it's a 1080p camera, while you can zoom in on a quadrant or part of the photo utilizing all possible pixels of the camera for that quadrant.
So much of smartphone cameras is in the software stack. AI processing, AF algorithms, all that image processing fun stuff is what gives each smartphone it's unique image look (and is what differentiates otherwise very similar phone cameras).
MP doesn't tell at all how good the quality of the image is, in general though the higher the MP the less quality you get, instead you have tons of noise and other artifacts
Jeff, your would be blown away if you investigated about the Gcam and what it does to compatible android devices...
I think I know what that banding might be its probably your lights when they brighten and dim from the ac voltage and the rolling shutter from the hawk eye is slow enough that you can see the brightening and dimming of your room whats making the rolling shutter slow idk
That's definitely a little bit. Apparently the yellow lines may be an artifact from some of the Pi's CSI pipeline... there's an option ArduCam mentioned that might be able to help.
Huh thats interesting I never guessed it would be the csi good to know
thanks for my friends, I wonder whether it is available on Pi Zero W and can be turned into manual focus.
want more like that
This would make an exceptional planetary camera. 😳 My current planetary cam has 2.5µ pixels and is insanely sharp. Just need one without the lens.
I teardowned my old samsung phone, and surprisingly, the cameras have the exact same kind of ribbon cable, and the text is kinda familar with the pi camera too
Hello Jeff ... Is it possible to use this cam simply as a magnifying glass with a maximum of 10x or 20x magnification?
If the driver supports it, there's a good reason for ePTZ out of the box -- you're limited to 1080p video, so... which 1080p region do you want?
3:09
Please do keep in mind though that the decreased quality of the comparison images is mostly due to the fact that the smartphones apply heavy luminance denoising that completely eats fine details, shoot in raw and the photos will be way more matched.
Hi Jeff,
a Nice project would be to use pi 5 and its 2 camera inputs,
and create a 3d recording setup and streaming it to some sort of app,
I am thinking on building a camera rig with a bunch of PIs for a photometry project. I wonder if PI5 would handle this camera or an updated version of this camera.
4:20 Hey, You've got the same speakers I do !