for a project im working on a very large interior of an old brewery.
At the moment im scanning just 1/20th of the overall surface. Due to the ceiling height I combined DSLR images with drone footage. When I first processed all 2.187 images at once it took the entire night just to align them and it created 20 components, though they are each very well aligned. I followed the guide “working with components” to export each component process it and merge them in a fresh project but processing spererately created just more component per component.
When I want to generate a dense pointcloud and mesh it takes longer then 48 hours. That is super odd, as usually all projects took only a few minutes to mere hours for large stuff.
I would like to ask for help on this one as it is exploding already. My rigg is not that bad actually, RTX 2080 OC, AMD Ryzen 7 2700 3.2ghz, 64GB Ram, Samsung EVO 980 PRO SSD, enough memory in the cache partition as well.
Any help is greatly appreciated. I could share the project, any detail etc.
The purpose is to build it into a pre-visualization in Unreal engine 5 as walk-through. Hence im looking for quite good detail.
what are your alignment and reconstruction settings? Can you describe your workflow more precise? Are you modelling using start button? It seems that you have merged the images into one component. Is this the result of the alignment or did you use the control points?
Is it necessary to compute this model in High details?
Here is a screenshot with all settings. Two days ago we finished the image collection. I used the start button at first but since it was soooo slow. (oddly slow) I started rather to go manually through the steps.
When I have multiple components after alignment. Does reality capture then process all components? or must I combine them? Because whenever I want to “merge components” nothing really changes. (must be something fundamentally that I did not do right)
For all the projects I did (usually smaller natural environments, small statues, small architecture detail work, somethings industrial) I honestly was ok with the start button process. I usually take the finished model and texture, project it onto a model with super clean topography and take the texture apart to create a pbr material. BUT this one is entirely different.
I want to take it as it is, drop it into UE5 and use the raw output.
so rather quick and dirty just to do lighting and pre-vizualisations. To get a look and feel of what could be done with the place.
In the screenshot you see one of the 27 rooms. They are super duper large. ;)
When you have more components you need to combine them to get one model. Are your components georeferenced? Did you used GCPs in your project? If so, you can use tool Merge georeferenced components only.
Did you also captured images for merging the rooms (images around doors or connection corridors)?
When you are talking about multiple components, is it about one room or all rooms?
I captured the portal part of the room with lots of shots but not the next rooms. All of the images have the same location but only arbitrary of course.
I find it so unusual that everything takes forever and ever and crashes usually. also that I can not combine the components is odd.
I think I may run out of memory? But each HDD has around 300GB space left.
Do you know how much is required for a scene like this? The reality capture temp file gets insanely huge.
Hi, so you just captured the portals, but not a connection between rooms? Then it could be possible, that it won’t be merged into one component, as there won’t be common parts… As I asked before, are there in the project ground control points with coordinates in some coordinate system? If so, it could be merge as georeferenced components. If not, you will need to capture another connection images.
Also I noticed, that your capturing style is not ideal, as it seems, that you were standing on one place and captured images from it and then you moved to another position (like pivot capturing):
This is not ideal for photogrammetry, as you can see on this comparison:
This long processing time can be related with the way how you captured your images. There is a short tutorial how to capture images for RealityCapture: https://www.youtube.com/watch?v=9e_NLp_FaUk&t=4s
In which step do you have these crashes? During model creation or alignment?
Yes, there could be a memory outage. For this it could achieve 1 TB of data (for all your models). Are you processing this on more computers?
I think we don’t have power and source to process such big project.
Hey, thanks for the detailed answer. I took the shooting style actually from one of the tutorials that was a french toomb or smth. The reason was mainly that it is an arc and I tried to follow the structure to have at least 60% overlap within the images.
The crash usually is in the model generation. I don’t have markers in the scene (should have done that).
BUT there is al ot of unique structure etc. that I could use for mering images or create controll points across images. I created one model sucessfully but there is no ceiling in the point cloud and no floor because I processed only one component for that. I have not figured out how to merge the component. they have so much overlap and actually created like 90% the same structure just across multiple components.
Maybe there is a threshhold or smth. that I may reduce to merge the components?
By ANY model with decent textures is ok for me. a
Regarding the portals: That worked at least well in alignment. I took a lot of connecting images so that it finds enough points to create the portal. In the pointcloud (of the other components) it is there together with the main chamber. If not, that is not a problem as it is not the priority now.
I need to showcase “this is how great it can be lit and used in UE based on the scan etc. etc. please unlock a budget so that we can do it properly”.
Since it is hard to get to the brewery cellar, I can not re-visit easily. This boils down to: How could I ideally merge the components? Do you have an advice or workaround on that? I’ll try to make more room on the SSD in the meantime and try to process it with less detail or so.
Actually:
It would be ok to export high density pointclouds as backup. But there again I have that issue that I need to bring the components together.
So for now, you have just one chamber captured? According to this it seems, that it is aligned correctly: - you have aligned all images. Are there more?
What are your reconstruction settings? This could be related with the disc space…
If it is just one chamber, you can send me the images, I will try to process them.
To merge the components, you can use control points, which will be place in both components.
I have been using Reality Capture for more than 3 years and am very interested in some complex scenes. When performing 3D reconstruction of complex scenes, if there are many rooms, it is recommended that you store the images of these rooms separately, and calculate the data separately.
RC provides a format for alignment files,after finish the images of all the rooms to be individually aligned, you can get the image alignment files for each room, such as room 1.rcalign, room 2.rcalign, room 3.rcalign, etc.,and then merge these alignment files in the new scene.
If you are willing to share your raw data I can help you try to calculate it, how many images you have ?
