RC failing at my image set

RC fails at my image set.
RC seems to use only like 10 images or so.

Here are the results of a 320 photos image-set (which should contain the entire high seat):

This is the dataset with the project files:
https://mab.to/mmexK8WCp

I have taken the images carefully with lots of overlaps, especially of the high seat itself at a far distance.
When I move in, I can’t take images from all sides, of course.

But the fact that RC creates only the seat doesn’t seem right to me.

The alignment settings are set to their defaults.

What am I doing wrong?

Thanks.

Some problems i found in your images:
ISO 200
All images with motion blur.
And it hard to say that images have “good overlap” .

With tuned settings it possible align about 285 from 310 images that is not bad for low quality images.

Thanks. Should I use ISO 100 instead?

Always strive to use the lowest ISO value your camera can do.

It minimizes noise.

To fight blurriness use either a tripod or a monopod to stabilize the camera where possible.
If you can connect a remote trigger to the camera, use it too. It prevents camera shake from pressing buttons on it.

I agree that SOME pictures are blurred, but most of them are fine.

How did you figure out the settings required to produce RC this better model?

Hi tw2016
in the images circled RED I can clearly see panoramic shots, and I think there are more of them… they are useless for photogramemtry…
the GREEN lines mean where you need to take images, say 30-50 images per round… you need to move closer and closer to the subject, not with so big jumps from each other, best if it is under 1/3 of the distance…

Thanks.

Is it allowed to discuss the outputs of other softwares compared to the RC output?

Hi tw2016
No, as this is a RealityCapture software forum page…

Hi Vladlen,

can you please show your aligment settings?
And how did you figure them out? I have tried so much, I couldn’t get your results.
Thanks!

Wishgranter wrote:

Hi tw2016

No as this is RealityCapture software forum page…

But how can we tell where RC need improvements if we don’t compare it to other softwares?
Or is RC a fixed system where we can’t suggest improvements?

Hi tw2016.

I think in this case it better just ask for new feature without any comparison.

But don’t worry, i also test your dataset in “other” software with the same result. So problem what i wrote (ISO noise, motion blur, DOF) and what Wishgranter draw (your cameras vs. how you should shoot such objects) for you is a main problem why you can’t have good result in most modern photogrammetry software.

As for settings:
I set bigger Feature per Image count, Reprojection error (around 4 or 5 pixels), Set image overlap to Low, And set bigger Preselector Features (about 1/2 of Feature count) in Advanced settings and Detector Sensitivity to High.
This was allow detect features that too weak (because of reason i wrote) and made “draft” traingulation of cameras. Later you can lower reprojection error delete all components except biggest, and run align again. This will allow refine alignment. But lower than 1px error you again lost most of camera positions. And this mean that mesh from such 1px error aligned cameras will have too much artifacts. Try imagine 1px in image can be 1-10 cm on real object, Every pixel - every 1-10 cm can be in random place.

Wishgranter wrote:

Hi tw2016
in the images circled RED I can clearly see panoramic shots, and I think there are more of them… they are useless for photogramemtry…
the GREEN lines mean where you need to take images, say 30-50 images per round… you need to move closer and closer to the subject, not with so big jumps from each other, best if it is under 1/3 of the distance…

I would counter, in a small way, the use of panoramic shots. If you’re trying to capture as much as possible, especially in a closed space, panoramics can be essential for a good workflow. I still hone in the way Wishgranter explains above on items of focus, which is what you absolutely need to for this example as you’re trying to get a single entity scanned, but if you were to get a whole section of the woods to explore before you hit that roost, I’d be using 15 to 20 degree rotations at about 8 locations around the eventual exploration box. Important stuff if your intent is to explore a large model for things like VR, not so much if you’re focusing on a handful of things only.

And while they really fill up unfocused parts of a large scene well, you need to be very careful on where you put them for best results. The first few iterations were horrible, but they really help me compose a large inside scene when I plan a shoot out.