New Xiaomi Mi Panoramic 360 Camera

Your camera is fine. That’s a. Samsung issue and there’s a solution for that. Give me a second to point out

1 Like

Join this group, a lot of knowledge and tips there

Log into Facebook | Facebook

1 Like

Nice, thank you a lot!

Small update.
After a few 2 seconds delay tests, I had to be satisfied with 3 seconds delay.
Usually when entering into shadow area from sunny area, camera was not ready to take next photo.
ISO 50 other settings auto (maybe will try with ISO 100)
3 seconds still look usable, although 2 seconds would be better.

2 seconds delay:

3 seconds delay:

1 Like

I’m also bought a Xiaomi for interval panorama shooting. My problem is that the images has poor image quality… The JPG image indeed has 6912x3456px, but when it is saved (both to JPG and) to RAW (it is very nice and good quality), the RAW has only 3456x1728 px. I could not obtain a 6912x3456 px RAW image… If I shooting JPG+RAW the interval between to photos (set to 2 sec) is about 8 seconds… If only the JPG is saved than is around 3 sec. I’m using as you suggested 128Gb Samsung Evo Plus U3 card.
The problem is that with the 6912x3456 px JPG image I can read the text (car’s plate number) just until 10 meters, so the farther texts will be unreadable…
Do you have a solution to have better resolution? The Gopro Fusion has less resolution, but maybe it’s more sharper then the Xiaomi… Did you tested the Fusion vs. Xiaomi?

Hi Jozsef, are you stitching the images on your phone or on a PC? I think the phone will stitch at a lower resolution if the phone is not powerful enough . I don’t use RAW capturing yet, so I may be unaware of other problems you may be meeting.

With that said, the resolution is not what people expect for a 16mpx photo, but remember that’s 16 megapixels over a full 360 degree field of view. My iphone gives me 12mpx but only a ~70degree field of view. This makes the detail a lot more concentrated.

When I survey with my Mi360, i keep my phone capturing too, so I have extra detail in one direction at least.

Hi Gness! I tried the stitching on computer, on phone. The resolution is not effected by the stitching on the phone.
The Xiaomi is marketed as a 24Mpixel camera with 6912x3456 pixel.
Based on my comparison the camera has only 3456x1728 pixel (since the RAW has only this resolution) and the JPG is artificially increased with interpolated pixels to double size. Based on my tests I can read the same text from the same distance in the same way from the double size JPG and the half size RAW. So in the JPG is no additional data.
I’m so sad to tell this, because I believed the true 6912x3456 pixel… and the Xiaomi is a good 360 camera which can be controlled externally by wire.

1 Like

In Nov 2017 @kaylesk posted a DNG (raw file) from the camera, in this thread, which I just pulled through exiftool. It reports the image is 6912x3456. The file size is 47.7 MB. I don’t have the camera myself but I think this deserves more investigation.

I didn’t know that. But I checked now the file that posted kaylesk and is 3456x1728 (half of the jpg resolution).
:frowning:

i just opened a raw image from this camera in photoshop and it says its 6912x3456
¯\(ツ)

heres an unprocessed jpg https://drive.google.com/open?id=1YyV1Kkn5cDYj4gKSftn2RtYjls6PRAbK
and a stiched https://drive.google.com/open?id=1zhSt714Z_W815kAK9HhaNSrl4pZ0pm57
both show 6912x3456
i havent shot in RAW

So we get to different results with the same file. Let’s try to pinpont the reason our results differ. Can I get you to post one of your raw files?

Here are four photos taken with the 6912x3456+RAW resolution setting. One pair shot from 125 cm and one from 100cm. The JPG has 6 Mb and 6912x3456 px and the DNG has 48Mb and 3456x1728 px like every other sample here on Mapillary forum.
https://drive.google.com/drive/folders/1MjW7KNfppH3p0Ol3t1yE1k5PXhUIqyYg

The problem is if you zoom in to see the letters and numbers printed on the box (with 28 pt Arial font - having 7.07 mm height - shot from 100cm and 125cm - it is the equivalent if you see a 7 cm height plate number from 10 meter and 125 meter) and compare it in the two images the JPG is only slightly better then the DNG. That is because the JPG is made by the manufacturer more brighter and with more contrast compared with the RAW. But since the resolution is double I expected a much more resolution on the JPG and I expected to read significantly further from the JPG. But the quality is quite the same. From the 125cm you can not read correctly the numbers nor from the JPG and nor from the DNG.

Other strange fact is that the JPG image has that “compression noise” everywhere, even on uniform colors, where normally the quality loss from the compression should not be visible.

This is why I believe that the JPG is extrapolated from the RAW to double adding additional pixels to get the 6912x3456 pixel size.

What do you think?

@Joze When I open IMG_20180625_134152_100cm.DNG in exiftool, it claims the file is 6912x3456. How do you determine that the file is smaller?

Always when I opened the DNG in the IrfanView it showed me 3456x1728 px. Now when you are telling me that you have in exiftool the full 6912x3456 px, I opened it into Photoshop here it showed me the same resolution that in exiftool for you: 6912x3456 px… So very probably the IrfanView software showed me a wrong resolution… I can not understand why…
I’ m sorry that I made this misleading.

I uploaded to the previous Google Drive folder also a print screen showing the IrfanView info about the DNG in this file: DNG_resolution_shown_in_IrfanView.jpg

@Jozsef Well it’s weird why IrfanView shows lower resolution. Even on info screen - http://temp.kayle.sk/irfanview.jpg

I replicated your small test and yes, letters are not readable, but you have 3456x3456 px for half sphere, only ~12 megapixels.

Here is sample from DSLR with fisheye http://temp.kayle.sk/fisheye.jpg (Samyang 8mm f/3.5 - Google Search)
Resolution is 3456x5184 and for full circle I made 5 shots (maybe 4 was ok, but 5 have larger overlapping) + up and down.
It’s 10-15m from guidepost. Height of those signs is 140mm and letters 15mm.

6912x3456px looks like a lot pixels, but I think it’s not so much for whole sphere.

Thanks for the example with the Samyang lens. I tried to build a street view solution, and therefore I would need many DSLR camera on the top of the car in the same rig. The problems with DSLR cameras is that most of them have mechanical shutter (and will be broken after 200-300.000 shots) and they are expensive since you must buy a lot of them to cover the 360x180 degree. The Samyang 8mm is a great lens but it has manual focus and I for Street View images I believe you must have auto focus lens.
Now I bought also a GoPro Fusion (which has less resolution: only 3104x3000 for a half sphere but the resulted images are very close (only a little bit worse) compared to Xiaomi Mijia Sphere 360, but it has the advantage of taking images at 0.5 sec (in fact is somewhere 0.75 sec after a small test - but I should test it again with a faster card) compared with the Sphere 360 (at which the minimum interval to set is 2 sec and the real interval is around 3 sec between the images).
Overall the Fusion is more suitable for taking Street View images, but you can not improve the GPS coordinate precision (you must use the built-in one), but for the Xiaomi you can trigger it from the pins externally and attach an RTK differential GPS to have more precise (sub-meter/a few decimeter) coordinates.
It might be a solution to have 2 or 3 Xiaomi cameras and to achieve 1.5 - 1 sec interval (for one Fusion price you can buy almost 3 Xiaomi camera), but I understood that the Xiaomi cameras have some optical/stitching problems and Mapillary does not recommend them because you can not measure it object accurately.
Do you have any feature extraction accuracy experience from the Xiaomi camera?

I have external intervalometer and with 2 seconds sometimes camera didn’t finish saving image in time. (Will try do more accurate 2 seconds test later)
Now I have 3 seconds interval - images.

1 second would be awesome. (for car, for walking 5-7 seconds is sufficient)

well… I’m looking at Reach M2 | RTK GNSS/GPS modules for high-precision mapping and thinking about this (or similar) solution. :slight_smile:

I have no information about this problem.

1 Like

I tried to open it with IrfanView (and plugins installed) and got the same low resolution as you.
But when I opened the same file with Rawtherapee (open source) it shows the full resolution. I have no idear why Irfanview shows a lower resolution, but the problem is not the file.

Now I am sure it has the full resolution!

Btw. Rawtherapee is a really nice rawconverter and editor, that has nice support for batch jobs - i.e. apply the same settings to a lot of images.