28mm model pics using turntable - Reconstruction Failed

[Brasstax]

I've taken pictures of a 28mm model using a turntable that allows photographs every 17 degrees. All of the pictures are taken "head-on". When I try to create a point cloud using the photographs, Zephyr either fails outright or can only identify a few points. I've gone through every photo and used the masking tool to try to help but I haven't had any more success. Any ideas on what I might be doing wrong?

The photos/mask files are here:

https://drive.google.com/drive/folders/1viOCsax1F8B67BjOuZdVXp_DuAYmWtqj?usp=sharing

[Kharn]

I am also looking to so a similiar thing, I have between 20-48 photos at my disposal and it only ever aligns 3 photos :\

[Roberto]

Hi Brasstax,
I had a look at the datasets and here are some suggestions:
- Take more photos to be sure to have enough overlap. Try to take 40-50 pictures.
- Cover as much area as possible with the subject (move the camera closer if necessary)
- You have to control the depth of field. I assume the camera is on a tripod with remote control shooting; you should raise the aperture to f/20 for example. With the current aperture value, the area in focus is quite narrow.

Hope this helps, keep trying you will learn with trial and error

[Brasstax]

I'll give those suggestions a try, thank you. A couple of follow-up questions:

- Would taking more photos "head-on" like the previous ones be sufficient or does the software require photos from different elevations as well (from partially above or below)?

- Does wide-angle distortion affect the software's ability to reconstruct?

[Roberto]

If you can do different orbits, it would be better, to reconstruct all the parts. I think the ideal case would be to do three orbits of 30 photos each. If you have to stay under 50 pictures, I would try two different angles with 25 photos each. You can freely use wide-angle lenses; if the dataset is good, then the radial distortion is compensated and handled without problems by the software.

[Kharn]

So is 30 pictures per orbit a good minimum? For example I have 48 photos, but its 4 seperate orbits at 12 photos a piece.

[Roberto]

Actually, the minimum depends on the subject and the camera used. As a rule of thumb, always remember that each point of the object surface should be viewed by at least three cameras to be reconstructed. 12x4 orbits can be fine, but I think you have to keep an offset between the orbits so that the software can stitch with an intermediate jump of a different orbit. If you instead take more photos in a single orbit, the acquisition is easier.

[Brasstax]

I tried 2 orbits of 25 photos each using a similar model that would not have as much of a problem with depth-of-field. The 'head-on' orbit was rejected entirely by the software but the '3/4 view' was fully reconstructed. The photos are at the link below:

https://drive.google.com/open?id=1GxKrqX6f6niYT5IgT8XakgESQv_EhOob

I'm unsure as to why one orbit would work better than the other. I've also noticed that this was the first time the photos were properly aligned as shot by the camera. Normally, if the reconstruction doesn't fail outright, the photos are all bunched up on top of each other. Does the software determine photo alignment based on photos that have 'passed'?

[Andrea Alessi]

Hi Brasstax,

a few suggestions:

- when using a turntable, always use masks. If you also use the "turntable mode" in masquerade, it will only take a few minutes. It took me 4 minutes to mask your dataset and it's always a good idea

- try to shoot in manual mode. Some of your images are shot at 1/250 sec, some other at 1/80 sec etc, which creates different lighting settings. While zephyr is robust to light changes, it's always better to have it as uniform as possible in your dataset, and since you are in a controlled environment you should aim to have the same lighting in all your images.

- try to use a lower iso. Always aim for ISO-100, use a tripod with remote and longer exposure to get better results. higher iso means higher noise (especially with consumer grade cameras), which leads to worse results in zephyr.

- If the two orbits are processed succesfully separately but zephyr is not able to create one single block, most likely you don't have enough vertical overlap. That's fine though, as you can easily fix this by simply taking a few shots that go from the bottom orbit to the top orbit. Or you can simply try increasing the parameters if processing time is not an issue. Even by masking i got 25/50 with close range/defaults, but moving to close range/deep succesfully allowed me to get 50/50. As you can see masking is essential in this turntable setup.



- below is my result from your dataset.I simply processed down the pipeline, so no editing at all (although i could remove some noise from the dense cloud i guess) If you can get sharper photos you will get much better geometry and much better texture. You can find some tips and tricks on photography for photogrammetry in this very short videotutorial.

Parameters used:

SfM: Close range/deep
MvS: Close range/defaults
Mesh: Close Range/defaults - sharp features
Mesh: additional photoconsistency run (clone & apply) with bad mesh/bad photos settings and fill holes filter, as the mesh had a lot of noise.
Textured mesh: defaults + blurriness check



- Masks, photos and zep file at the following URL (will be deleted in a couple of months or let me know if you'd like me to delete it now) :

https://shared.3dflow.net/index.php/s/aLzL0a1waMVbV1F

- finally, consider also doing a third orbit from below (or even better, rotate the object, as you are going to mask anyways) so you can also have view from the bottom side (most of the time you will need more than the 50 images zephyr free offers though, unless the subject has a very easy geometry like the dismal souvenir dataset)

[Brasstax]

Thanks for the in-depth analysis Andrea! It was immensely helpful.