Capturing from a Ricoh Theta (360 image) fails

Development Asset Creation
,
1

Hi All,

I tried to capture a small space with a 360 camera (images exported to 6 cube faces), and it completely failed. But I want to know why.

I took 38 shots, removed the bottom image (which had the tripod), which resulted in 190 cube faces roughly like this:

However, when running through RC, I get 23 components (usually with 2-6 cams), and the alignment is all over the place.

Is this because the image is square? Because the distortion correction is unusual? To my eye, the distortion actually looks very good.

At the end of the day my goal is simply to use a 360 camera for scanning (b/c of the high amount of information a 360 image has).

2

Hi breravin

  1. i see a WHITE walls without any “features” that could be used in alignment

  2. img looks a bit noisy even here in web view.

  3. can try capture it in “photogrammetry friendly” environment ?

3

This was just an example image which had nice clean lines for distortion comparison. My real capture has a lot more information (but it is potentially sensitive, so I cannot post it here).

It includes a desk, computer, keyboard, monitor, chair, boxes, and cabinets. I also have alignment markers placed all around the room. Still, nothing generates in Reality Capture.

This is a fairly straightforward 3m x 2.5m space that I’m looking to recreate. It’s not particularly “unfriendly” to photogrammetry.

Even with the lower quality images, I would expect there to be a lot more alignment.

4

From what I’ve seen of the camera you’re going to be pushing it a bit in terms of image quality. Probably not helped by reprojecting the stitched image to cube faces. If you can get the raw image from each lens that should work better. Just paint over the tripod with white. Either way I’d consider using a different camera.

For a 20mp circular fisheye image I find the mesh gets noisy after 2m. While you can clearly see further than that, the image resolution beyond that distance isn’t very useful for photogrammetry. Try and think of an effective image circle of 120° per lens and a max working distance of 1m when considering camera positions and you should get enough overlap… and use 3 different camera heights.

Changing the sequence of images may also help… Sort the images so that they get listed by face and rename them eg. left01, left02… right01, right02… etc or if you’re using the raw image from each lens,split them into front and back sequences.

5

Thanks Ben, good thoughts. Somewhat confirming my suspicions a bit. I’m “glad” both of you feel it’s a quality issue (and not a “distortion” issue). That at least gives me something to go on.

Regarding using the raw image - do you mean the actual 360 image? Will RC actually handle that? I haven’t tried that myself, but I will try on Monday when I’m back in the office.

6

Definitely not the 360° image. This has been stitched and there is some parallax between the two source images. Can you get the images from each lens separately?

7

I’m not sure on the Ricoh Theta - I’ll look into it!

8

Modeling images from a 360 degree camera